From ce0c90eccda50c13e52fc52e65acf22257b21720 Mon Sep 17 00:00:00 2001 From: mhjensen Date: Tue, 16 Oct 2018 12:21:57 +0200 Subject: [PATCH] Update on neural network slides --- doc/pub/NeuralNet/html/._NeuralNet-bs000.html | 132 +- doc/pub/NeuralNet/html/._NeuralNet-bs001.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs002.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs003.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs004.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs005.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs006.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs007.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs008.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs009.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs010.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs011.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs012.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs013.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs014.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs015.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs016.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs017.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs018.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs019.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs020.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs021.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs022.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs023.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs024.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs025.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs026.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs027.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs028.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs029.html | 132 +- doc/pub/NeuralNet/html/._NeuralNet-bs030.html | 136 +- doc/pub/NeuralNet/html/._NeuralNet-bs031.html | 130 +- doc/pub/NeuralNet/html/._NeuralNet-bs032.html | 133 +- doc/pub/NeuralNet/html/._NeuralNet-bs033.html | 184 ++- doc/pub/NeuralNet/html/._NeuralNet-bs034.html | 231 +-- doc/pub/NeuralNet/html/._NeuralNet-bs035.html | 238 +-- doc/pub/NeuralNet/html/._NeuralNet-bs036.html | 193 +-- doc/pub/NeuralNet/html/._NeuralNet-bs037.html | 197 +-- doc/pub/NeuralNet/html/._NeuralNet-bs038.html | 187 +-- doc/pub/NeuralNet/html/._NeuralNet-bs039.html | 169 ++- doc/pub/NeuralNet/html/._NeuralNet-bs040.html | 202 +-- doc/pub/NeuralNet/html/._NeuralNet-bs041.html | 208 ++- doc/pub/NeuralNet/html/._NeuralNet-bs042.html | 172 ++- doc/pub/NeuralNet/html/._NeuralNet-bs043.html | 177 ++- doc/pub/NeuralNet/html/._NeuralNet-bs044.html | 256 ++-- doc/pub/NeuralNet/html/._NeuralNet-bs045.html | 240 ++- doc/pub/NeuralNet/html/._NeuralNet-bs046.html | 243 +-- doc/pub/NeuralNet/html/._NeuralNet-bs047.html | 250 +-- doc/pub/NeuralNet/html/._NeuralNet-bs048.html | 176 +-- doc/pub/NeuralNet/html/._NeuralNet-bs049.html | 183 ++- doc/pub/NeuralNet/html/._NeuralNet-bs050.html | 189 +-- doc/pub/NeuralNet/html/._NeuralNet-bs051.html | 191 +-- doc/pub/NeuralNet/html/._NeuralNet-bs052.html | 175 ++- doc/pub/NeuralNet/html/._NeuralNet-bs053.html | 174 +-- doc/pub/NeuralNet/html/._NeuralNet-bs054.html | 217 ++- doc/pub/NeuralNet/html/._NeuralNet-bs055.html | 320 ++-- doc/pub/NeuralNet/html/._NeuralNet-bs056.html | 323 ++-- doc/pub/NeuralNet/html/._NeuralNet-bs057.html | 203 ++- doc/pub/NeuralNet/html/._NeuralNet-bs058.html | 235 ++- doc/pub/NeuralNet/html/._NeuralNet-bs059.html | 164 +- doc/pub/NeuralNet/html/._NeuralNet-bs060.html | 171 ++- doc/pub/NeuralNet/html/._NeuralNet-bs061.html | 175 ++- doc/pub/NeuralNet/html/._NeuralNet-bs062.html | 183 ++- doc/pub/NeuralNet/html/._NeuralNet-bs063.html | 162 +- doc/pub/NeuralNet/html/._NeuralNet-bs064.html | 340 +++++ doc/pub/NeuralNet/html/._NeuralNet-bs065.html | 321 ++++ doc/pub/NeuralNet/html/NeuralNet-bs.html | 132 +- doc/pub/NeuralNet/html/NeuralNet-reveal.html | 304 ++-- .../NeuralNet/html/NeuralNet-solarized.html | 350 +++-- doc/pub/NeuralNet/html/NeuralNet.html | 350 +++-- doc/pub/NeuralNet/ipynb/NeuralNet.ipynb | 1343 ++++------------- .../ipynb/ipynb-NeuralNet-src.tar.gz | Bin 210 -> 210 bytes doc/pub/NeuralNet/pdf/NeuralNet-minted.pdf | Bin 467117 -> 469564 bytes doc/src/NeuralNet/NeuralNet.do.txt | 223 ++- 74 files changed, 7348 insertions(+), 6706 deletions(-) create mode 100644 doc/pub/NeuralNet/html/._NeuralNet-bs064.html create mode 100644 doc/pub/NeuralNet/html/._NeuralNet-bs065.html diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs000.html b/doc/pub/NeuralNet/html/._NeuralNet-bs000.html index e3971cad4..aeabd57ad 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs000.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs000.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -272,7 +276,7 @@ MathJax.Hub.Config({
    [2] Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University

    -

    Oct 12, 2018

    +

    Oct 16, 2018


    @@ -296,7 +300,7 @@ MathJax.Hub.Config({

  • 9
  • 10
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs001.html b/doc/pub/NeuralNet/html/._NeuralNet-bs001.html index 54cd21351..a370554d4 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs001.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs001.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -281,7 +285,7 @@ a weight variable.
  • 10
  • 11
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs002.html b/doc/pub/NeuralNet/html/._NeuralNet-bs002.html index 9bcaea022..4a9148147 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs002.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs002.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -331,7 +335,7 @@ humanities to life science and medicine.
  • 11
  • 12
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs003.html b/doc/pub/NeuralNet/html/._NeuralNet-bs003.html index 4ed1e3131..dc57c22b0 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs003.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs003.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -296,7 +300,7 @@ methods we discussed earlier.
  • 12
  • 13
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs004.html b/doc/pub/NeuralNet/html/._NeuralNet-bs004.html index f71e20e69..567102e7a 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs004.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs004.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -288,7 +292,7 @@ to all nodes in the subsequent layer, making this a so-called
  • 13
  • 14
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs005.html b/doc/pub/NeuralNet/html/._NeuralNet-bs005.html index 183d3814f..873e374bd 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs005.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs005.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -297,7 +301,7 @@ recognition.
  • 14
  • 15
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs006.html b/doc/pub/NeuralNet/html/._NeuralNet-bs006.html index 3af21dbc7..3f1994444 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs006.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs006.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -289,7 +293,7 @@ especially well-suited for handwriting and speech recognition.
  • 15
  • 16
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs007.html b/doc/pub/NeuralNet/html/._NeuralNet-bs007.html index 3de076525..a31b5dce4 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs007.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs007.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -290,7 +294,7 @@ type of NN due the unusual activation functions.
  • 16
  • 17
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs008.html b/doc/pub/NeuralNet/html/._NeuralNet-bs008.html index 67c9494d4..2fae7fb2c 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs008.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs008.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -288,7 +292,7 @@ Such networks are often called multilayer perceptrons (MLPs).
  • 17
  • 18
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs009.html b/doc/pub/NeuralNet/html/._NeuralNet-bs009.html index 92c158ee4..d6be8593d 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs009.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs009.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -293,7 +297,7 @@ as to not restrict the range of output values.
  • 18
  • 19
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs010.html b/doc/pub/NeuralNet/html/._NeuralNet-bs010.html index 6c356cc44..682cd9df8 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs010.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs010.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -294,7 +298,7 @@ of the outputs of all neurons in the previous layer.
  • 19
  • 20
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs011.html b/doc/pub/NeuralNet/html/._NeuralNet-bs011.html index aa1667957..0eea9b850 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs011.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs011.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -323,7 +327,7 @@ is obtained.
  • 20
  • 21
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs012.html b/doc/pub/NeuralNet/html/._NeuralNet-bs012.html index 037a4eca5..490a37900 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs012.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs012.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -305,7 +309,7 @@ $$
  • 21
  • 22
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs013.html b/doc/pub/NeuralNet/html/._NeuralNet-bs013.html index 9b6a78abc..598a73d5d 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs013.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs013.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -296,7 +300,7 @@ variables are the input values \( x_n \).
  • 22
  • 23
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs014.html b/doc/pub/NeuralNet/html/._NeuralNet-bs014.html index 7be0680d6..89f79e50f 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs014.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs014.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -305,7 +309,7 @@ flexibility of a neural network.
  • 23
  • 24
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs015.html b/doc/pub/NeuralNet/html/._NeuralNet-bs015.html index 28b61905e..c00812ea2 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs015.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs015.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -315,7 +319,7 @@ $$
  • 24
  • 25
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs016.html b/doc/pub/NeuralNet/html/._NeuralNet-bs016.html index e01825146..662103b8a 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs016.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs016.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -299,7 +303,7 @@ used as input to the activation functions. For each operation
  • 25
  • 26
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs017.html b/doc/pub/NeuralNet/html/._NeuralNet-bs017.html index d0ee040d6..4a4434ae1 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs017.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs017.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -293,7 +297,7 @@ for a FFNN to fulfill the universal approximation theorem
  • 26
  • 27
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs018.html b/doc/pub/NeuralNet/html/._NeuralNet-bs018.html index 7d73e4600..73bbfd2a3 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs018.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs018.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -301,7 +305,7 @@ $$
  • 27
  • 28
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs019.html b/doc/pub/NeuralNet/html/._NeuralNet-bs019.html index 0bea91398..0bd555817 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs019.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs019.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -363,7 +367,7 @@ plt.show()
  • 28
  • 29
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs020.html b/doc/pub/NeuralNet/html/._NeuralNet-bs020.html index a1714bf72..41fba85a8 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs020.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs020.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -314,7 +318,7 @@ like logistic regression or linear regression and their modifications on the oth
  • 29
  • 30
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs021.html b/doc/pub/NeuralNet/html/._NeuralNet-bs021.html index 0fe5f24de..ff17bac08 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs021.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs021.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -304,7 +308,7 @@ the potential of being universal approximators.
  • 30
  • 31
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs022.html b/doc/pub/NeuralNet/html/._NeuralNet-bs022.html index 3d052b9eb..588834449 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs022.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs022.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -310,7 +314,7 @@ classes.
  • 31
  • 32
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs023.html b/doc/pub/NeuralNet/html/._NeuralNet-bs023.html index d866cefe0..00261c1b3 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs023.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs023.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -314,7 +318,7 @@ $$
  • 32
  • 33
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs024.html b/doc/pub/NeuralNet/html/._NeuralNet-bs024.html index 345fb6270..c3cbc66b1 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs024.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs024.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -298,7 +302,7 @@ $$
  • 33
  • 34
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs025.html b/doc/pub/NeuralNet/html/._NeuralNet-bs025.html index e7175d481..a8ebde711 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs025.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs025.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -301,7 +305,7 @@ $$
  • 34
  • 35
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs026.html b/doc/pub/NeuralNet/html/._NeuralNet-bs026.html index 703180894..823d0411c 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs026.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs026.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -326,7 +330,7 @@ $$
  • 35
  • 36
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs027.html b/doc/pub/NeuralNet/html/._NeuralNet-bs027.html index 8c0d23e2d..e998bcea1 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs027.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs027.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -293,7 +297,7 @@ That is, the error \( \delta_j^L \) is exactly equal to the rate of change of th
  • 36
  • 37
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs028.html b/doc/pub/NeuralNet/html/._NeuralNet-bs028.html index cb6c8133f..181635ec9 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs028.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs028.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -338,7 +342,7 @@ one \( L-1 \) in terms of the errors in the final output layer.
  • 37
  • 38
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs029.html b/doc/pub/NeuralNet/html/._NeuralNet-bs029.html index 841bd0bd5..a77376e91 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs029.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs029.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -272,7 +276,7 @@ $$ z_j^{l+1} = \sum_{i=1}^{M_{l}}w_{ij}^{l+1}a_j^{l}+b_j^{l+1}, $$ -we obtain +with \( M_l \) being the number of nodes in layer \( l \), we obtain $$ \delta_j^l =\sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l), $$ @@ -308,7 +312,7 @@ We are now ready to set up the algorithm for back propagation and learning the w
  • 38
  • 39
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs030.html b/doc/pub/NeuralNet/html/._NeuralNet-bs030.html index 520ffdf16..34cf226ff 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs030.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs030.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -299,7 +303,7 @@ $$

    Then we compute the back propagate error for each \( l=L-1,L-2,\dots,2 \) as $$ -\delta_j^l =\sum_k \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l). +\delta_j^l = \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l). $$ @@ -309,14 +313,14 @@ $$

    -Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,dots,2 \) and update the weights and biases according to the rules +Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,\dots,2 \) and update the weights and biases according to the rules $$ w_{jk}^l\leftarrow = w_{jk}^l- \eta \delta_j^la_k^{l-1}, $$ $$ -b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^L}, +b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^l}=b_j^l-\eta \delta_j^l, $$

    @@ -352,7 +356,7 @@ Here it is convenient to use stochastic gradient descent (see the examples below
  • 39
  • 40
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs031.html b/doc/pub/NeuralNet/html/._NeuralNet-bs031.html index 65c9b4686..e930b5526 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs031.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs031.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -314,7 +318,7 @@ of our network.
  • 40
  • 41
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs032.html b/doc/pub/NeuralNet/html/._NeuralNet-bs032.html index 792c3e430..418007b57 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs032.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs032.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -302,14 +306,13 @@ $$ Again we take the negative log-likelihood to define our cost function: $$ -\mathcal{C}(\hat{\theta}) = - \ln P(\mathcal{D} \mid \hat{\theta}). +\mathcal{C}(\hat{\theta}) = - \log{P(\mathcal{D} \mid \hat{\theta})}. $$ See the logistic regression lectures for a full definition of the cost function.

    The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before! -We leave it as an exercise in project 2 to derive these equations.

    @@ -337,7 +340,7 @@ We leave it as an exercise in project 2 to derive these equations.

  • 41
  • 42
  • ...
    • -
  • 64
    • +
  • 66
  • »
    • diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs033.html b/doc/pub/NeuralNet/html/._NeuralNet-bs033.html index a71e6cd84..c949a9d21 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs033.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs033.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -251,22 +255,54 @@ MathJax.Hub.Config({

     

     

     

    - + -

    Developing a code for doing neural networks with back propagation

    +

    Example: binary classification problem

    -One can identify a set of key steps when using neural networks to solve supervised learning problems: +As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters \( \beta \) as +$$ +\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right), +$$ -

      -
    1. Collect and pre-process data
      2. -
    2. Define model and architecture
      3. -
    3. Choose cost function and optimizer
      4. -
    4. Train the model
      5. -
    5. Evaluate model performance on test data
      6. -
    6. Adjust hyperparameters (if necessary, network architecture)
      7. -
    +where we had defined the logistic (sigmoid) function +$$ +p(y_i =1\vert x_i,\hat{\beta})=\frac{\exp{(\beta_0+\beta_1 x_i)}}{1+\exp{(\beta_0+\beta_1 x_i)}}, +$$ +and +$$ +p(y_i =0\vert x_i,\hat{\beta})=1-p(y_i =1\vert x_i,\hat{\beta}). +$$ + +The parameters \( \hat{\beta} \) were defined using a minimization method like gradient descent or Newton-Raphson's method. + +

    +Now we replace \( x_i \) with the activation \( z_i^l \) for a given layer \( l \) and the outputs as \( y_i=a_i^l=f(z_i^l) \), with \( z_i^l \) now being a function of the weights \( w_{ij}^l \) and biases \( b_i^l \). +We have then +$$ +a_i^l = y_i = \frac{\exp{(z_i^l)}}{1+\exp{(z_i^l)}}, +$$ + +with +$$ +z_i^l = \sum_{j}w_{ij}^l a_j^{l-1}+b_i^l, +$$ + +where the superscript \( l-1 \) indicates that these are the outputs from layer \( l-1 \). +Our cost function at the final layer \( l=L \) is now +$$ +\mathcal{C}(\hat{W}) = - \sum_{i=1}^n \left(t_i\log{a_i^L}+(i-t_i)\log{(1-a_i^L)}\right), +$$ + +where we have defined the targets \( t_i \). The derivatives of the cost function with respect to the output \( a_i^L \) are then easily calculated and we get +$$ +\frac{\partial \mathcal{C}(\hat{W})}{\partial a_i^L} = \frac{a_i^L-t_i}{a_i^L(1-a_i^L)}. +$$ + +In case we use another activation function than the logistic one, we need to evaluate other derivatives. + +

    diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs034.html b/doc/pub/NeuralNet/html/._NeuralNet-bs034.html index 37dbc6a56..c5af01771 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs034.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs034.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -251,101 +255,22 @@ MathJax.Hub.Config({

     

     

     

    - + -

    Collect and pre-process data

    +

    Developing a code for doing neural networks with back propagation

    -Here we will be using the MNIST dataset, which is readily available through the scikit-learn -package. You may also find it for example here. -The MNIST (Modified National Institute of Standards and Technology) database is a large database -of handwritten digits that is commonly used for training various image processing systems. -The MNIST dataset consists of 70 000 images of size 28x28 pixels, each labeled from 0 to 9. -The scikit-learn dataset we will use consists of a selection of 1797 images of size \( 8\times 8 \) collected and processed from this database. +One can identify a set of key steps when using neural networks to solve supervised learning problems: -

    -To feed data into a feed-forward neural network we need to represent -the inputs as a feature matrix \( X = (n_{inputs}, n_{features}) \). Each -row represents an input, in this case a handwritten digit, and -each column represents a feature, in this case a pixel. The -correct answers, also known as labels or targets are -represented as a 1D array of integers -\( Y = (n_{inputs}) = (5, 3, 1, 8,...) \). +

      +
    1. Collect and pre-process data
      2. +
    2. Define model and architecture
      3. +
    3. Choose cost function and optimizer
      4. +
    4. Train the model
      5. +
    5. Evaluate model performance on test data
      6. +
    6. Adjust hyperparameters (if necessary, network architecture)
      7. +
    -

    -As an example, say we want to build a neural network using supervised learning to predict Body-Mass Index (BMI) from -measurements of height (in m) -and weight (in kg). If we have measurements of 5 people the feature matrix could be for example: - -$$ X = \begin{bmatrix} -1.85 & 81\\ -1.71 & 65\\ -1.95 & 103\\ -1.55 & 42\\ -1.63 & 56 -\end{bmatrix} ,$$ - -

    -and the targets would be: - -$$ Y = (23.7, 22.2, 27.1, 17.5, 21.1) $$ - -

    -Since each input image is a 2D matrix, we need to flatten the image -(i.e. "unravel" the 2D matrix into a 1D array) to turn the data into a -feature matrix. This means we lose all spatial information in the -image, such as locality and translational invariance. More complicated -architectures such as Convolutional Neural Networks can take advantage -of such information, and are most commonly applied when analyzing -images. - -

    - - -

    # import necessary packages
-import numpy as np
-import matplotlib.pyplot as plt
-from sklearn import datasets
-
-
-# ensure the same random numbers appear every time
-np.random.seed(0)
-
-# display images in notebook
-%matplotlib inline
-plt.rcParams['figure.figsize'] = (12,12)
-
-
-# download MNIST dataset
-digits = datasets.load_digits()
-
-# define inputs and labels
-inputs = digits.images
-labels = digits.target
-
-print("inputs = (n_inputs, pixel_width, pixel_height) = " + str(inputs.shape))
-print("labels = (n_inputs) = " + str(labels.shape))
-
-
-# flatten the image
-# the value -1 means dimension is inferred from the remaining dimensions: 8x8 = 64
-n_inputs = len(inputs)
-inputs = inputs.reshape(n_inputs, -1)
-print("X = (n_inputs, n_features) = " + str(inputs.shape))
-
-
-# choose some random images to display
-indices = np.arange(n_inputs)
-random_indices = np.random.choice(indices, size=5)
-
-for i, image in enumerate(digits.images[random_indices]):
-    plt.subplot(1, 5, i+1)
-    plt.axis('off')
-    plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
-    plt.title("Label: %d" % digits.target[random_indices[i]])
-plt.show()
-
    -

    diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs035.html b/doc/pub/NeuralNet/html/._NeuralNet-bs035.html index 58cbb1712..c031fd78d 100644 --- a/doc/pub/NeuralNet/html/._NeuralNet-bs035.html +++ b/doc/pub/NeuralNet/html/._NeuralNet-bs035.html @@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source None, '___sec30'), ('Defining the cost function', 2, None, '___sec31'), + ('Example: binary classification problem', 2, None, '___sec32'), ('Developing a code for doing neural networks with back ' 'propagation', 2, None, - '___sec32'), - ('Collect and pre-process data', 2, None, '___sec33'), - ('Train and test datasets', 2, None, '___sec34'), - ('Define model and architecture', 2, None, '___sec35'), - ('Layers', 2, None, '___sec36'), - ('Weights and biases', 2, None, '___sec37'), - ('Feed-forward pass', 2, None, '___sec38'), - ('Matrix multiplications', 2, None, '___sec39'), - ('Choose cost function and optimizer', 2, None, '___sec40'), - ('Optimizing the cost function', 2, None, '___sec41'), - ('Regularization', 2, None, '___sec42'), - ('Matrix multiplication', 2, None, '___sec43'), - ('Improving performance', 2, None, '___sec44'), - ('Full object-oriented implementation', 2, None, '___sec45'), - ('Evaluate model performance on test data', 2, None, '___sec46'), - ('Adjust hyperparameters', 2, None, '___sec47'), - ('Visualization', 2, None, '___sec48'), - ('scikit-learn implementation', 2, None, '___sec49'), - ('Visualization', 2, None, '___sec50'), + '___sec33'), + ('Collect and pre-process data', 2, None, '___sec34'), + ('Train and test datasets', 2, None, '___sec35'), + ('Define model and architecture', 2, None, '___sec36'), + ('Layers', 2, None, '___sec37'), + ('Weights and biases', 2, None, '___sec38'), + ('Feed-forward pass', 2, None, '___sec39'), + ('Matrix multiplications', 2, None, '___sec40'), + ('Choose cost function and optimizer', 2, None, '___sec41'), + ('Optimizing the cost function', 2, None, '___sec42'), + ('Regularization', 2, None, '___sec43'), + ('Matrix multiplication', 2, None, '___sec44'), + ('Improving performance', 2, None, '___sec45'), + ('Full object-oriented implementation', 2, None, '___sec46'), + ('Evaluate model performance on test data', 2, None, '___sec47'), + ('Adjust hyperparameters', 2, None, '___sec48'), + ('Visualization', 2, None, '___sec49'), + ('scikit-learn implementation', 2, None, '___sec50'), + ('Visualization', 2, None, '___sec51'), ('Building neural networks in Tensorflow and Keras', 2, None, - '___sec51'), - ('Tensorflow', 2, None, '___sec52'), - ('Collect and pre-process data', 2, None, '___sec53'), - ('Using TensorFlow backend', 2, None, '___sec54'), - ('Optimizing and using gradient descent', 2, None, '___sec55'), - ('Using Keras', 2, None, '___sec56'), - ('Which activation function should I use?', 2, None, '___sec57'), + '___sec52'), + ('Tensorflow', 2, None, '___sec53'), + ('Collect and pre-process data', 2, None, '___sec54'), + ('Using TensorFlow backend', 2, None, '___sec55'), + ('Optimizing and using gradient descent', 2, None, '___sec56'), + ('Using Keras', 2, None, '___sec57'), + ('Which activation function should I use?', 2, None, '___sec58'), ('Is the Logistic activation function (Sigmoid) our choice?', 2, None, - '___sec58'), - ('The derivative of the Logistic funtion', 2, None, '___sec59'), - ('The RELU function family', 2, None, '___sec60'), + '___sec59'), + ('The derivative of the Logistic funtion', 2, None, '___sec60'), + ('The RELU function family', 2, None, '___sec61'), + ('Which activation function should we use?', 2, None, '___sec62'), ('A top-down perspective on Neural networks', 2, None, - '___sec61'), + '___sec63'), ('Limitations of supervised learning with deep networks', 2, None, - '___sec62')]} + '___sec64')]} end of tocinfo --> @@ -207,37 +209,39 @@ MathJax.Hub.Config({
  • Setting up the Back propagation algorithm
  • Setting up a Multi-layer perceptron model for classification
  • Defining the cost function
    • -
  • Developing a code for doing neural networks with back propagation
    • -
  • Collect and pre-process data
    • -
  • Train and test datasets
    • -
  • Define model and architecture
    • -
  • Layers
    • -
  • Weights and biases
    • -
  • Feed-forward pass
    • -
  • Matrix multiplications
    • -
  • Choose cost function and optimizer
    • -
  • Optimizing the cost function
    • -
  • Regularization
    • -
  • Matrix multiplication
    • -
  • Improving performance
    • -
  • Full object-oriented implementation
    • -
  • Evaluate model performance on test data
    • -
  • Adjust hyperparameters
    • -
  • Visualization
    • -
  • scikit-learn implementation
    • -
  • Visualization
    • -
  • Building neural networks in Tensorflow and Keras
    • -
  • Tensorflow
    • -
  • Collect and pre-process data
    • -
  • Using TensorFlow backend
    • -
  • Optimizing and using gradient descent
    • -
  • Using Keras
    • -
  • Which activation function should I use?
    • -
  • Is the Logistic activation function (Sigmoid) our choice?
    • -
  • The derivative of the Logistic funtion
    • -
  • The RELU function family
    • -
  • A top-down perspective on Neural networks
    • -
  • Limitations of supervised learning with deep networks
    • +
  • Example: binary classification problem
    • +
  • Developing a code for doing neural networks with back propagation
    • +
  • Collect and pre-process data
    • +
  • Train and test datasets
    • +
  • Define model and architecture
    • +
  • Layers
    • +
  • Weights and biases
    • +
  • Feed-forward pass
    • +
  • Matrix multiplications
    • +
  • Choose cost function and optimizer
    • +
  • Optimizing the cost function
    • +
  • Regularization
    • +
  • Matrix multiplication
    • +
  • Improving performance
    • +
  • Full object-oriented implementation
    • +
  • Evaluate model performance on test data
    • +
  • Adjust hyperparameters
    • +
  • Visualization
    • +
  • scikit-learn implementation
    • +
  • Visualization
    • +
  • Building neural networks in Tensorflow and Keras
    • +
  • Tensorflow
    • +
  • Collect and pre-process data
    • +
  • Using TensorFlow backend
    • +
  • Optimizing and using gradient descent
    • +
  • Using Keras
    • +
  • Which activation function should I use?
    • +
  • Is the Logistic activation function (Sigmoid) our choice?
    • +
  • The derivative of the Logistic funtion
    • +
  • The RELU function family
    • +
  • Which activation function should we use?
    • +
  • A top-down perspective on Neural networks
    • +
  • Limitations of supervised learning with deep networks
    • @@ -253,51 +257,97 @@ MathJax.Hub.Config({ -

    Train and test datasets

    +

    Collect and pre-process data

    -Performing analysis before partitioning the dataset is a major error, that can lead to incorrect conclusions. +Here we will be using the MNIST dataset, which is readily available through the scikit-learn +package. You may also find it for example here. +The MNIST (Modified National Institute of Standards and Technology) database is a large database +of handwritten digits that is commonly used for training various image processing systems. +The MNIST dataset consists of 70 000 images of size 28x28 pixels, each labeled from 0 to 9. +The scikit-learn dataset we will use consists of a selection of 1797 images of size \( 8\times 8 \) collected and processed from this database.

    -We will reserve \( 80 \% \) of our dataset for training and \( 20 \% \) for testing. +To feed data into a feed-forward neural network we need to represent +the inputs as a feature matrix \( X = (n_{inputs}, n_{features}) \). Each +row represents an input, in this case a handwritten digit, and +each column represents a feature, in this case a pixel. The +correct answers, also known as labels or targets are +represented as a 1D array of integers +\( Y = (n_{inputs}) = (5, 3, 1, 8,...) \).

    -It is important that the train and test datasets are drawn randomly from our dataset, to ensure -no bias in the sampling. -Say you are taking measurements of weather data to predict the weather in the coming 5 days. -You don't want to train your model on measurements taken from the hours 00.00 to 12.00, and then test it on data -collected from 12.00 to 24.00. +As an example, say we want to build a neural network using supervised learning to predict Body-Mass Index (BMI) from +measurements of height (in m) +and weight (in kg). If we have measurements of 5 people the feature matrix could be for example: + +$$ X = \begin{bmatrix} +1.85 & 81\\ +1.71 & 65\\ +1.95 & 103\\ +1.55 & 42\\ +1.63 & 56 +\end{bmatrix} ,$$ + +

    +and the targets would be: + +$$ Y = (23.7, 22.2, 27.1, 17.5, 21.1) $$ + +

    +Since each input image is a 2D matrix, we need to flatten the image +(i.e. "unravel" the 2D matrix into a 1D array) to turn the data into a +feature matrix. This means we lose all spatial information in the +image, such as locality and translational invariance. More complicated +architectures such as Convolutional Neural Networks can take advantage +of such information, and are most commonly applied when analyzing +images.

    -

    from sklearn.model_selection import train_test_split
+
    # import necessary packages
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn import datasets
 
-# one-liner from scikit-learn library
-train_size = 0.8
-test_size = 1 - train_size
-X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=train_size,
-                                                    test_size=test_size)
 
-# equivalently in numpy
-def train_test_split_numpy(inputs, labels, train_size, test_size):
-    n_inputs = len(inputs)
-    inputs_shuffled = inputs.copy()
-    labels_shuffled = labels.copy()
-    
-    np.random.shuffle(inputs_shuffled)
-    np.random.shuffle(labels_shuffled)
-    
-    train_end = int(n_inputs*train_size)
-    X_train, X_test = inputs_shuffled[:train_end], inputs_shuffled[train_end:]
-    Y_train, Y_test = labels_shuffled[:train_end], labels_shuffled[train_end:]
-    
-    return X_train, X_test, Y_train, Y_test
+# ensure the same random numbers appear every time
+np.random.seed(0)
 
-#X_train, X_test, Y_train, Y_test = train_test_split_numpy(inputs, labels, train_size, test_size)
+# display images in notebook
+%matplotlib inline
+plt.rcParams['figure.figsize'] = (12,12)
 
-print("Number of training images: " + str(len(X_train)))
-print("Number of test images: " + str(len(X_test)))
+
+# download MNIST dataset
+digits = datasets.load_digits()
+
+# define inputs and labels
+inputs = digits.images
+labels = digits.target
+
+print("inputs = (n_inputs, pixel_width, pixel_height) = " + str(inputs.shape))
+print("labels = (n_inputs) = " + str(labels.shape))
+
+
+# flatten the image
+# the value -1 means dimension is inferred from the remaining dimensions: 8x8 = 64
+n_inputs = len(inputs)
+inputs = inputs.reshape(n_inputs, -1)
+print("X = (n_inputs, n_features) = " + str(inputs.shape))
+
+
+# choose some random images to display
+indices = np.arange(n_inputs)
+random_indices = np.random.choice(indices, size=5)
+
+for i, image in enumerate(digits.images[random_indices]):
+    plt.subplot(1, 5, i+1)
+    plt.axis('off')
+    plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
+    plt.title("Label: %d" % digits.target[random_indices[i]])
+plt.show()
 
    
 
    
 
    
@@ -325,7 +375,7 @@ X_train, X_test, Y_train, Y_test = train_tes
   
  • 44
    • 
   
  • 45
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs036.html b/doc/pub/NeuralNet/html/._NeuralNet-bs036.html
index 3d5fec962..b82e20c62 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs036.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs036.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,49 +257,52 @@ MathJax.Hub.Config({
 
 
 
-
    Define model and architecture 
    
+
    Train and test datasets 
    
 
 
    
-Our simple feed-forward neural network will consist of an input layer, a single hidden layer and an output layer. The activation \( y \) of each neuron is a weighted sum of inputs, passed through an activation function. In case of the simple perceptron model we have 
-
-$$ z = \sum_{i=1}^n w_i a_i ,$$
-
-$$ y = f(z) ,$$
+Performing analysis before partitioning the dataset is a major error, that can lead to incorrect conclusions.
 
 
    
-where \( f \) is the activation function, \( a_i \) represents input from neuron \( i \) in the preceding layer
-and \( w_i \) is the weight to input \( i \).  
-The activation of the neurons in the input layer is just the features (e.g. a pixel value).
+We will reserve \( 80 \% \) of our dataset for training and \( 20 \% \) for testing.
 
 
    
-The simplest activation function for a neuron is the Heaviside function:
-
-$$ f(z) = 
-\begin{cases}
-1,  &  z > 0\\
-0,  & \text{otherwise}
-\end{cases}
-$$
+It is important that the train and test datasets are drawn randomly from our dataset, to ensure
+no bias in the sampling.  
+Say you are taking measurements of weather data to predict the weather in the coming 5 days.
+You don't want to train your model on measurements taken from the hours 00.00 to 12.00, and then test it on data
+collected from 12.00 to 24.00.
 
 
    
-A feed-forward neural network with this activation is known as a perceptron.  
-For a binary classifier (i.e. two classes, 0 or 1, dog or not-dog) we can also use this in our output layer.  
-This activation can be generalized to \( k \) classes (using e.g. the one-against-all strategy), 
-and we call these architectures multiclass perceptrons.
 
-
    
-However, it is now common to use the terms Single Layer Perceptron (SLP) (1 hidden layer) and  
-Multilayer Perceptron (MLP) (2 or more hidden layers) to refer to feed-forward neural networks with any activation function.
+
+
    from sklearn.model_selection import train_test_split
 
-
    
-Typical choices for activation functions include the sigmoid function, hyperbolic tangent, and Rectified Linear Unit (ReLU).  
-We will be using the sigmoid function \( \sigma(x) \):  
+# one-liner from scikit-learn library
+train_size = 0.8
+test_size = 1 - train_size
+X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=train_size,
+                                                    test_size=test_size)
 
-$$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
+# equivalently in numpy
+def train_test_split_numpy(inputs, labels, train_size, test_size):
+    n_inputs = len(inputs)
+    inputs_shuffled = inputs.copy()
+    labels_shuffled = labels.copy()
+    
+    np.random.shuffle(inputs_shuffled)
+    np.random.shuffle(labels_shuffled)
+    
+    train_end = int(n_inputs*train_size)
+    X_train, X_test = inputs_shuffled[:train_end], inputs_shuffled[train_end:]
+    Y_train, Y_test = labels_shuffled[:train_end], labels_shuffled[train_end:]
+    
+    return X_train, X_test, Y_train, Y_test
 
-
    
-which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.
+#X_train, X_test, Y_train, Y_test = train_test_split_numpy(inputs, labels, train_size, test_size)
 
+print("Number of training images: " + str(len(X_train)))
+print("Number of test images: " + str(len(X_test)))
+
    
 
    
 
    
 
@@ -322,7 +329,7 @@ which is inspired by probability theory (see logistic regression) and was most c
   
  • 45
    • 
   
  • 46
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs037.html b/doc/pub/NeuralNet/html/._NeuralNet-bs037.html
index 815f47594..74ef76c31 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs037.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs037.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,49 +255,50 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Layers 
    
-
-
      
-
    •  Input
      •  
-
    
-
-Since each input image has 8x8 = 64 pixels or features, we have an input layer of 64 neurons.  
-
-
      
-
    •  Hidden layer
      • 
-
    
-
-We will use 50 neurons in the hidden layer receiving input from the neurons in the input layer.  
-Since each neuron in the hidden layer is connected to the 64 inputs we have 64x50 = 3200 weights to the hidden layer.  
-
-
      
-
    •  Output
      • 
-
    
-
-If we were building a binary classifier, it would be sufficient with a single neuron in the output layer,
-which could output 0 or 1 according to the Heaviside function. This would be an example of a hard classifier, meaning it outputs the class of the input directly. However, if we are dealing with noisy data it is often beneficial to use a soft classifier, which outputs the probability of being in class 0 or 1.
+
    Define model and architecture 
    
 
 
    
-For a soft binary classifier, we could use a single neuron and interpret the output as either being the probability of being in class 0 or the probability of being in class 1. Alternatively we could use 2 neurons, and interpret each neuron as the probability of being in each class.
+Our simple feed-forward neural network will consist of an input layer, a single hidden layer and an output layer. The activation \( y \) of each neuron is a weighted sum of inputs, passed through an activation function. In case of the simple perceptron model we have 
+
+$$ z = \sum_{i=1}^n w_i a_i ,$$
+
+$$ y = f(z) ,$$
 
 
    
-Since we are doing multiclass classification, with 10 categories, it is natural to use 10 neurons in the output layer. We number the neurons \( j = 0,1,...,9 \). The activation of each output neuron \( j \) will be according to the softmax function:  
-
-$$ P(\text{class \( j \)} \mid \text{input \( \hat{a} \)}) = \frac{\exp{(\hat{a}^T \hat{w}_j)}}
-{\sum_{c=0}^{9} \exp{(\hat{a}^T \hat{w}_c)}} ,$$
+where \( f \) is the activation function, \( a_i \) represents input from neuron \( i \) in the preceding layer
+and \( w_i \) is the weight to input \( i \).  
+The activation of the neurons in the input layer is just the features (e.g. a pixel value).
 
 
    
-i.e. each neuron \( j \) outputs the probability of being in class \( j \) given an input from the hidden layer \( \hat{a} \), with \( \hat{w}_j \) the weights of neuron \( j \) to the inputs.  
-The denominator is a normalization factor to ensure the outputs (probabilities) sum up to 1.  
-The exponent is just the weighted sum of inputs as before:  
+The simplest activation function for a neuron is the Heaviside function:
 
-$$ z_j = \sum_{i=1}^n w_ {ij} a_i+b_j.$$
+$$ f(z) = 
+\begin{cases}
+1,  &  z > 0\\
+0,  & \text{otherwise}
+\end{cases}
+$$
 
 
    
-Since each neuron in the output layer is connected to the 50 inputs from the hidden layer we have 50x10 = 500
-weights to the output layer.
+A feed-forward neural network with this activation is known as a perceptron.  
+For a binary classifier (i.e. two classes, 0 or 1, dog or not-dog) we can also use this in our output layer.  
+This activation can be generalized to \( k \) classes (using e.g. the one-against-all strategy), 
+and we call these architectures multiclass perceptrons.
+
+
    
+However, it is now common to use the terms Single Layer Perceptron (SLP) (1 hidden layer) and  
+Multilayer Perceptron (MLP) (2 or more hidden layers) to refer to feed-forward neural networks with any activation function.
+
+
    
+Typical choices for activation functions include the sigmoid function, hyperbolic tangent, and Rectified Linear Unit (ReLU).  
+We will be using the sigmoid function \( \sigma(x) \):  
+
+$$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
+
+
    
+which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.
 
 
    
 
    
@@ -321,7 +326,7 @@ weights to the output layer.
   
  • 46
    • 
   
  • 47
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs038.html b/doc/pub/NeuralNet/html/._NeuralNet-bs038.html
index 499facaf5..fe1f53db9 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs038.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs038.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,39 +257,48 @@ MathJax.Hub.Config({
 
 
 
-
    Weights and biases 
    
+
    Layers 
    
+
+
      
+
    •  Input
      •  
+
    
+
+Since each input image has 8x8 = 64 pixels or features, we have an input layer of 64 neurons.  
+
+
      
+
    •  Hidden layer
      • 
+
    
+
+We will use 50 neurons in the hidden layer receiving input from the neurons in the input layer.  
+Since each neuron in the hidden layer is connected to the 64 inputs we have 64x50 = 3200 weights to the hidden layer.  
+
+
      
+
    •  Output
      • 
+
    
+
+If we were building a binary classifier, it would be sufficient with a single neuron in the output layer,
+which could output 0 or 1 according to the Heaviside function. This would be an example of a hard classifier, meaning it outputs the class of the input directly. However, if we are dealing with noisy data it is often beneficial to use a soft classifier, which outputs the probability of being in class 0 or 1.
 
 
    
-Typically weights are initialized with small values distributed around zero, drawn from a uniform
-or normal distribution. Setting all weights to zero means all neurons give the same output, making the network useless.
+For a soft binary classifier, we could use a single neuron and interpret the output as either being the probability of being in class 0 or the probability of being in class 1. Alternatively we could use 2 neurons, and interpret each neuron as the probability of being in each class.
 
 
    
-Adding a bias value to the weighted sum of inputs allows the neural network to represent a greater range
-of values. Without it, any input with the value 0 will be mapped to zero (before being passed through the activation). The bias unit has an output of 1, and a weight to each neuron \( j \), \( b_j \):  
+Since we are doing multiclass classification, with 10 categories, it is natural to use 10 neurons in the output layer. We number the neurons \( j = 0,1,...,9 \). The activation of each output neuron \( j \) will be according to the softmax function:  
 
-$$ z_j = \sum_{i=1}^n w_ {ij} a_i + b_j.$$
+$$ P(\text{class \( j \)} \mid \text{input \( \hat{a} \)}) = \frac{\exp{(\hat{a}^T \hat{w}_j)}}
+{\sum_{c=0}^{9} \exp{(\hat{a}^T \hat{w}_c)}} ,$$
 
 
    
-The bias weights \( \hat{b} \) are often initialized to zero, but a small value like \( 0.01 \) ensures all neurons have some output which can be backpropagated in the first training cycle.
+i.e. each neuron \( j \) outputs the probability of being in class \( j \) given an input from the hidden layer \( \hat{a} \), with \( \hat{w}_j \) the weights of neuron \( j \) to the inputs.  
+The denominator is a normalization factor to ensure the outputs (probabilities) sum up to 1.  
+The exponent is just the weighted sum of inputs as before:  
+
+$$ z_j = \sum_{i=1}^n w_ {ij} a_i+b_j.$$
+
 
    
+Since each neuron in the output layer is connected to the 50 inputs from the hidden layer we have 50x10 = 500
+weights to the output layer.
 
-
-
    # building our neural network
-
-n_inputs, n_features = X_train.shape
-n_hidden_neurons = 50
-n_categories = 10
-
-# we make the weights normally distributed using numpy.random.randn
-
-# weights and bias in the hidden layer
-hidden_weights = np.random.randn(n_features, n_hidden_neurons)
-hidden_bias = np.zeros(n_hidden_neurons) + 0.01
-
-# weights and bias in the output layer
-output_weights = np.random.randn(n_hidden_neurons, n_categories)
-output_bias = np.zeros(n_categories) + 0.01
-
    
 
    
 
    
 
@@ -312,7 +325,7 @@ output_bias = np47
   
  • 48
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs039.html b/doc/pub/NeuralNet/html/._NeuralNet-bs039.html
index c5fc1e55a..2f21e4b4f 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs039.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs039.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,32 +255,41 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Feed-forward pass 
    
+
    Weights and biases 
    
 
 
    
-Denote \( F \) the number of features, \( H \) the number of hidden neurons and \( C \) the number of categories.  
-For each input image we calculate a weighted sum of input features (pixel values) to each neuron \( j \) in the hidden layer \( l \):  
-
-$$ z_{j}^{l} = \sum_{i=1}^{F} w_{ij}^{l} x_i + b_{j}^{l},$$
+Typically weights are initialized with small values distributed around zero, drawn from a uniform
+or normal distribution. Setting all weights to zero means all neurons give the same output, making the network useless.
 
 
    
-this is then passed through our activation function  
+Adding a bias value to the weighted sum of inputs allows the neural network to represent a greater range
+of values. Without it, any input with the value 0 will be mapped to zero (before being passed through the activation). The bias unit has an output of 1, and a weight to each neuron \( j \), \( b_j \):  
 
-$$ a_{j}^{l} = f(z_{j}^{l}) .$$
+$$ z_j = \sum_{i=1}^n w_ {ij} a_i + b_j.$$
 
 
    
-We calculate a weighted sum of inputs (activations in the hidden layer) to each neuron \( j \) in the output layer:  
-
-$$ z_{j}^{L} = \sum_{i=1}^{H} w_{ij}^{L} a_{i}^{l} + b_{j}^{L}.$$
-
+The bias weights \( \hat{b} \) are often initialized to zero, but a small value like \( 0.01 \) ensures all neurons have some output which can be backpropagated in the first training cycle.
 
    
-Finally we calculate the output of neuron \( j \) in the output layer using the softmax function:  
 
-$$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
-{\sum_{c=0}^{C-1} \exp{(z_c^{L})}} .$$
+
+
    # building our neural network
 
+n_inputs, n_features = X_train.shape
+n_hidden_neurons = 50
+n_categories = 10
+
+# we make the weights normally distributed using numpy.random.randn
+
+# weights and bias in the hidden layer
+hidden_weights = np.random.randn(n_features, n_hidden_neurons)
+hidden_bias = np.zeros(n_hidden_neurons) + 0.01
+
+# weights and bias in the output layer
+output_weights = np.random.randn(n_hidden_neurons, n_categories)
+output_bias = np.zeros(n_categories) + 0.01
+
    
 
    
 
    
 
@@ -303,7 +316,7 @@ $$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
   
  • 48
    • 
   
  • 49
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs040.html b/doc/pub/NeuralNet/html/._NeuralNet-bs040.html
index 079fbbb01..ee0727ce9 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs040.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs040.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,80 +255,32 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Matrix multiplications 
    
+
    Feed-forward pass 
    
 
 
    
-Since our data has the dimensions \( X = (n_{inputs}, n_{features}) \) and our weights to the hidden
-layer have the dimensions  
-\( W_{hidden} = (n_{features}, n_{hidden}) \),
-we can easily feed the network all our training data in one go by taking the matrix product  
+Denote \( F \) the number of features, \( H \) the number of hidden neurons and \( C \) the number of categories.  
+For each input image we calculate a weighted sum of input features (pixel values) to each neuron \( j \) in the hidden layer \( l \):  
 
-$$ X W^{h} = (n_{inputs}, n_{hidden}),$$
+$$ z_{j}^{l} = \sum_{i=1}^{F} w_{ij}^{l} x_i + b_{j}^{l},$$
 
 
    
-and obtain a matrix that holds the weighted sum of inputs to the hidden layer
-for each input image and each hidden neuron.    
-We also add the bias to obtain a matrix of weighted sums to the hidden layer \( Z^{h} \):  
+this is then passed through our activation function  
 
-$$ \hat{z}^{l} = \hat{X} \hat{W}^{l} + \hat{b}^{l} ,$$
+$$ a_{j}^{l} = f(z_{j}^{l}) .$$
 
 
    
-meaning the same bias (1D array with size equal number of hidden neurons) is added to each input image.  
-This is then passed through the activation:  
+We calculate a weighted sum of inputs (activations in the hidden layer) to each neuron \( j \) in the output layer:  
 
-$$ \hat{a}^{l} = f(\hat{z}^l) .$$
+$$ z_{j}^{L} = \sum_{i=1}^{H} w_{ij}^{L} a_{i}^{l} + b_{j}^{L}.$$
 
 
    
-This is fed to the output layer:  
+Finally we calculate the output of neuron \( j \) in the output layer using the softmax function:  
 
-$$ \hat{z}^{L} = \hat{a}^{L} \hat{W}^{L} + \hat{b}^{L} .$$
+$$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
+{\sum_{c=0}^{C-1} \exp{(z_c^{L})}} .$$
 
-
    
-Finally we receive our output values for each image and each category by passing it through the softmax function:  
-
-$$ output = softmax (\hat{z}^{L}) = (n_{inputs}, n_{categories}) .$$
-
-
    
-
-
-
    # setup the feed-forward pass, subscript h = hidden layer
-
-def sigmoid(x):
-    return 1/(1 + np.exp(-x))
-
-def feed_forward(X):
-    # weighted sum of inputs to the hidden layer
-    z_h = np.matmul(X, hidden_weights) + hidden_bias
-    # activation in the hidden layer
-    a_h = sigmoid(z_h)
-    
-    # weighted sum of inputs to the output layer
-    z_o = np.matmul(a_h, output_weights) + output_bias
-    # softmax output
-    # axis 0 holds each input and axis 1 the probabilities of each category
-    exp_term = np.exp(z_o)
-    probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
-    
-    return probabilities
-
-probabilities = feed_forward(X_train)
-print("probabilities = (n_inputs, n_categories) = " + str(probabilities.shape))
-print("probability that image 0 is in category 0,1,2,...,9 = \n" + str(probabilities[0]))
-print("probabilities sum up to: " + str(probabilities[0].sum()))
-print()
-
-# we obtain a prediction by taking the class with the highest likelihood
-def predict(X):
-    probabilities = feed_forward(X)
-    return np.argmax(probabilities, axis=1)
-
-predictions = predict(X_train)
-print("predictions = (n_inputs) = " + str(predictions.shape))
-print("prediction for image 0: " + str(predictions[0]))
-print("correct label for image 0: " + str(Y_train[0]))
-
    
 
    
 
    
 
@@ -351,7 +307,7 @@ predictions = predict(X_train)
   
  • 49
    • 
   
  • 50
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs041.html b/doc/pub/NeuralNet/html/._NeuralNet-bs041.html
index 728bc868a..d01d443c9 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs041.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs041.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,36 +255,80 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Choose cost function and optimizer 
    
+
    Matrix multiplications 
    
 
 
    
-To measure how well our neural network is doing we need to introduce a cost function.  
-We will call the function that gives the error of a single sample output the loss function, and the function
-that gives the total error of our network across all samples the cost function.
-A typical choice for multiclass classification is the cross-entropy loss, also known as the negative log likelihood.
+Since our data has the dimensions \( X = (n_{inputs}, n_{features}) \) and our weights to the hidden
+layer have the dimensions  
+\( W_{hidden} = (n_{features}, n_{hidden}) \),
+we can easily feed the network all our training data in one go by taking the matrix product  
+
+$$ X W^{h} = (n_{inputs}, n_{hidden}),$$
 
 
    
-In multiclass classification it is common to treat each integer label as a so called one-hot vector:  
+and obtain a matrix that holds the weighted sum of inputs to the hidden layer
+for each input image and each hidden neuron.    
+We also add the bias to obtain a matrix of weighted sums to the hidden layer \( Z^{h} \):  
 
-$$ y = 5 \quad \rightarrow \quad \hat{y} = (0, 0, 0, 0, 0, 1, 0, 0, 0, 0) ,$$  
-
-$$ y = 1 \quad \rightarrow \quad \hat{y} = (0, 1, 0, 0, 0, 0, 0, 0, 0, 0) ,$$
+$$ \hat{z}^{l} = \hat{X} \hat{W}^{l} + \hat{b}^{l} ,$$
 
 
    
-i.e. a binary bit string of length \( C \), where \( C = 10 \) is the number of classes in the MNIST dataset.
+meaning the same bias (1D array with size equal number of hidden neurons) is added to each input image.  
+This is then passed through the activation:  
+
+$$ \hat{a}^{l} = f(\hat{z}^l) .$$
 
 
    
-Let \( y_{ic} \) denote the \( c \)-th component of the \( i \)-th one-hot vector.  
-We define the cost function \( \mathcal{C} \) as a sum over the cross-entropy loss for each point \( \hat{x}_i \) in the dataset.
+This is fed to the output layer:  
+
+$$ \hat{z}^{L} = \hat{a}^{L} \hat{W}^{L} + \hat{b}^{L} .$$
 
 
    
-In the one-hot representation only one of the terms in the loss function is non-zero, namely the
-probability of the correct category \( c' \)  
-(i.e. the category \( c' \) such that \( y_{ic'} = 1 \)). This means that the cross entropy loss only punishes you for how wrong
-you got the correct label. The probability of category \( c \) is given by the softmax function. The vector \( \hat{\theta} \) represents the parameters of our network, i.e. all the weights and biases.
+Finally we receive our output values for each image and each category by passing it through the softmax function:  
 
+$$ output = softmax (\hat{z}^{L}) = (n_{inputs}, n_{categories}) .$$
+
+
    
+
+
+
    # setup the feed-forward pass, subscript h = hidden layer
+
+def sigmoid(x):
+    return 1/(1 + np.exp(-x))
+
+def feed_forward(X):
+    # weighted sum of inputs to the hidden layer
+    z_h = np.matmul(X, hidden_weights) + hidden_bias
+    # activation in the hidden layer
+    a_h = sigmoid(z_h)
+    
+    # weighted sum of inputs to the output layer
+    z_o = np.matmul(a_h, output_weights) + output_bias
+    # softmax output
+    # axis 0 holds each input and axis 1 the probabilities of each category
+    exp_term = np.exp(z_o)
+    probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
+    
+    return probabilities
+
+probabilities = feed_forward(X_train)
+print("probabilities = (n_inputs, n_categories) = " + str(probabilities.shape))
+print("probability that image 0 is in category 0,1,2,...,9 = \n" + str(probabilities[0]))
+print("probabilities sum up to: " + str(probabilities[0].sum()))
+print()
+
+# we obtain a prediction by taking the class with the highest likelihood
+def predict(X):
+    probabilities = feed_forward(X)
+    return np.argmax(probabilities, axis=1)
+
+predictions = predict(X_train)
+print("predictions = (n_inputs) = " + str(predictions.shape))
+print("prediction for image 0: " + str(predictions[0]))
+print("correct label for image 0: " + str(Y_train[0]))
+
    
 
    
 
    
 
@@ -307,7 +355,7 @@ you got the correct label. The probability of category \( c \) is given by the s
   
  • 50
    • 
   
  • 51
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs042.html b/doc/pub/NeuralNet/html/._NeuralNet-bs042.html
index c57b99214..c51a3bae6 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs042.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs042.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,41 +257,35 @@ MathJax.Hub.Config({
 
 
 
-
    Optimizing the cost function 
    
+
    Choose cost function and optimizer 
    
 
 
    
-The network is trained by finding the weights and biases that minimize the cost function. One of the most widely used classes of methods is gradient descent and its generalizations. The idea behind gradient descent
-is simply to adjust the weights in the direction where the gradient of the cost function is large and negative. This ensures we flow toward a local minimum of the cost function.  
-Each parameter \( \theta \) is iteratively adjusted according to the rule  
-
-$$ \theta_{i+1} = \theta_i - \eta \nabla \mathcal{C}(\theta_i) ,$$
+To measure how well our neural network is doing we need to introduce a cost function.  
+We will call the function that gives the error of a single sample output the loss function, and the function
+that gives the total error of our network across all samples the cost function.
+A typical choice for multiclass classification is the cross-entropy loss, also known as the negative log likelihood.
 
 
    
-where \( \eta \) is known as the learning rate, which controls how big a step we take towards the minimum.  
-This update can be repeated for any number of iterations, or until we are satisfied with the result.
+In multiclass classification it is common to treat each integer label as a so called one-hot vector:  
+
+$$ y = 5 \quad \rightarrow \quad \hat{y} = (0, 0, 0, 0, 0, 1, 0, 0, 0, 0) ,$$  
+
+$$ y = 1 \quad \rightarrow \quad \hat{y} = (0, 1, 0, 0, 0, 0, 0, 0, 0, 0) ,$$
 
 
    
-A simple and effective improvement is a variant called Batch Gradient Descent.  
-Instead of calculating the gradient on the whole dataset, we calculate an approximation of the gradient
-on a subset of the data called a minibatch.  
-If there are \( N \) data points and we have a minibatch size of \( M \), the total number of batches
-is \( N/M \).  
-We denote each minibatch \( B_k \), with \( k = 1, 2,...,N/M \). The gradient then becomes:  
-
-$$ \nabla \mathcal{C}(\theta) = \frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}_i(\theta) \quad \rightarrow \quad
-\frac{1}{M} \sum_{i \in B_k} \nabla \mathcal{L}_i(\theta) ,$$
+i.e. a binary bit string of length \( C \), where \( C = 10 \) is the number of classes in the MNIST dataset.
 
 
    
-i.e. instead of averaging the loss over the entire dataset, we average over a minibatch.
+Let \( y_{ic} \) denote the \( c \)-th component of the \( i \)-th one-hot vector.  
+We define the cost function \( \mathcal{C} \) as a sum over the cross-entropy loss for each point \( \hat{x}_i \) in the dataset.
 
 
    
-This has two important benefits:  
-
-
      
-
    1.  Introducing stochasticity decreases the chance that the algorithm becomes stuck in a local minima.
      2.   
-
    2.  It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
      3.   
-
    
+In the one-hot representation only one of the terms in the loss function is non-zero, namely the
+probability of the correct category \( c' \)  
+(i.e. the category \( c' \) such that \( y_{ic'} = 1 \)). This means that the cross entropy loss only punishes you for how wrong
+you got the correct label. The probability of category \( c \) is given by the softmax function. The vector \( \hat{\theta} \) represents the parameters of our network, i.e. all the weights and biases.
 
+
    
 
    
 
 
      
@@ -313,7 +311,7 @@ This has two important benefits:
   
    • 51
      • 
   
    • 52
      • 
   
    • ...
      • 
-  
    • 64
      • 
+  
    • 66
      • 
   
    • »
      • 
 
    
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs043.html b/doc/pub/NeuralNet/html/._NeuralNet-bs043.html
index e252abadf..c8a2b2793 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs043.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs043.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,37 +255,44 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Regularization 
    
+
    Optimizing the cost function 
    
 
 
    
-It is common to add an extra term to the cost function, proportional
-to the size of the weights.  This is equivalent to constraining the
-size of the weights, so that they do not grow out of control.
-Constraining the size of the weights means that the weights cannot
-grow arbitrarily large to fit the training data, and in this way
-reduces overfitting.
+The network is trained by finding the weights and biases that minimize the cost function. One of the most widely used classes of methods is gradient descent and its generalizations. The idea behind gradient descent
+is simply to adjust the weights in the direction where the gradient of the cost function is large and negative. This ensures we flow toward a local minimum of the cost function.  
+Each parameter \( \theta \) is iteratively adjusted according to the rule  
+
+$$ \theta_{i+1} = \theta_i - \eta \nabla \mathcal{C}(\theta_i) ,$$
 
 
    
-We will measure the size of the weights using the so called L2-norm, meaning our cost function becomes:  
+where \( \eta \) is known as the learning rate, which controls how big a step we take towards the minimum.  
+This update can be repeated for any number of iterations, or until we are satisfied with the result.
+
+
    
+A simple and effective improvement is a variant called Batch Gradient Descent.  
+Instead of calculating the gradient on the whole dataset, we calculate an approximation of the gradient
+on a subset of the data called a minibatch.  
+If there are \( N \) data points and we have a minibatch size of \( M \), the total number of batches
+is \( N/M \).  
+We denote each minibatch \( B_k \), with \( k = 1, 2,...,N/M \). The gradient then becomes:  
 
 $$ \nabla \mathcal{C}(\theta) = \frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}_i(\theta) \quad \rightarrow \quad
-\frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}_i(\theta) + \lambda \lvert \lvert \hat{w} \rvert \rvert_2^2 
-= \frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}(\theta) + \lambda \sum_{ij} w_{ij}^2,$$
+\frac{1}{M} \sum_{i \in B_k} \nabla \mathcal{L}_i(\theta) ,$$
 
 
    
-i.e. we sum up all the weights squared. The factor \( \lambda \) is known as a regularization parameter.
+i.e. instead of averaging the loss over the entire dataset, we average over a minibatch.
 
 
    
-In order to train the model, we need to calculate the derivative of
-the cost function with respect to every bias and weight in the
-network.  In total our network has \( (64 + 1)\times 50=3250 \) weights in
-the hidden layer and \( (50 + 1)\times 10=510 \) weights to the output
-layer (\( +1 \) for the bias), and the gradient must be calculated for
-every parameter.  We use the backpropagation algorithm discussed
-above. This is a clever use of the chain rule that allows us to
-calculate the gradient efficently.
+This has two important benefits:  
+
+
      
+
    1.  Introducing stochasticity decreases the chance that the algorithm becomes stuck in a local minima.
      2.   
+
    2.  It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
      3.   
+
    
+
+The various optmization  methods, with codes and algorithms,  are discussed in our lectures on Gradient descent approaches.
 
 
    
 
    
@@ -309,7 +320,7 @@ calculate the gradient efficently.
   
  • 52
    • 
   
  • 53
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs044.html b/doc/pub/NeuralNet/html/._NeuralNet-bs044.html
index c9412a9bc..734d0f29b 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs044.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs044.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,122 +255,38 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Matrix  multiplication 
    
+
    Regularization 
    
 
 
    
-To more efficently train our network these equations are implemented using matrix operations.  
-The error in the output layer is calculated simply as  
-
-$$ \delta_L = \hat{y} - y = (n_{inputs}, n_{categories}) .$$
+It is common to add an extra term to the cost function, proportional
+to the size of the weights.  This is equivalent to constraining the
+size of the weights, so that they do not grow out of control.
+Constraining the size of the weights means that the weights cannot
+grow arbitrarily large to fit the training data, and in this way
+reduces overfitting.
 
 
    
-The gradient for the output weights is calculated as  
+We will measure the size of the weights using the so called L2-norm, meaning our cost function becomes:  
 
-$$ \nabla W_{L} = \hat{a}^T \delta_L   = (n_{hidden}, n_{categories}) ,$$
+$$ \nabla \mathcal{C}(\theta) = \frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}_i(\theta) \quad \rightarrow \quad
+\frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}_i(\theta) + \lambda \lvert \lvert \hat{w} \rvert \rvert_2^2 
+= \frac{1}{N} \sum_{i=1}^N \nabla \mathcal{L}(\theta) + \lambda \sum_{ij} w_{ij}^2,$$
 
 
    
-where \( \hat{a} = (n_{inputs}, n_{hidden}) \). This simply means that we are summing up the gradients for each input.  
-Since we are going backwards we have to transpose the activation matrix.
+i.e. we sum up all the weights squared. The factor \( \lambda \) is known as a regularization parameter.
 
 
    
-The gradient with respect to the output bias is then  
+In order to train the model, we need to calculate the derivative of
+the cost function with respect to every bias and weight in the
+network.  In total our network has \( (64 + 1)\times 50=3250 \) weights in
+the hidden layer and \( (50 + 1)\times 10=510 \) weights to the output
+layer (\( +1 \) for the bias), and the gradient must be calculated for
+every parameter.  We use the backpropagation algorithm discussed
+above. This is a clever use of the chain rule that allows us to
+calculate the gradient efficently.
 
-$$ \nabla \hat{b}_{L} = \sum_{i=1}^{n_{inputs}} \delta_L = (n_{categories}) .$$
-
-
    
-The error in the hidden layer is  
-
-$$ \Delta_h = \delta_L W_{L}^T \circ f'(z_{h}) = \delta_L W_{L}^T \circ a_{h} \circ (1 - a_{h}) = (n_{inputs}, n_{hidden}) ,$$
-
-
    
-where \( f'(a_{h}) \) is the derivative of the activation in the hidden layer. The matrix products mean
-that we are summing up the products for each neuron in the output layer. The symbol \( \circ \) denotes
-the Hadamard product, meaning element-wise multiplication.
-
-
    
-This again gives us the gradients in the hidden layer:  
-
-$$ \nabla W_{h} = X^T \delta_h = (n_{features}, n_{hidden}) ,$$  
-
-$$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
-
-
    
-
-
-
    # to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our model
-from sklearn.metrics import accuracy_score
-
-# one-hot in numpy
-def to_categorical_numpy(integer_vector):
-    n_inputs = len(integer_vector)
-    n_categories = np.max(integer_vector) + 1
-    onehot_vector = np.zeros((n_inputs, n_categories))
-    onehot_vector[range(n_inputs), integer_vector] = 1
-    
-    return onehot_vector
-
-#Y_train_onehot, Y_test_onehot = to_categorical(Y_train), to_categorical(Y_test)
-Y_train_onehot, Y_test_onehot = to_categorical_numpy(Y_train), to_categorical_numpy(Y_test)
-
-def feed_forward_train(X):
-    # weighted sum of inputs to the hidden layer
-    z_h = np.matmul(X, hidden_weights) + hidden_bias
-    # activation in the hidden layer
-    a_h = sigmoid(z_h)
-    
-    # weighted sum of inputs to the output layer
-    z_o = np.matmul(a_h, output_weights) + output_bias
-    # softmax output
-    # axis 0 holds each input and axis 1 the probabilities of each category
-    exp_term = np.exp(z_o)
-    probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
-    
-    # for backpropagation need activations in hidden and output layers
-    return a_h, probabilities
-
-def backpropagation(X, Y):
-    a_h, probabilities = feed_forward_train(X)
-    
-    # error in the output layer
-    error_output = probabilities - Y
-    # error in the hidden layer
-    error_hidden = np.matmul(error_output, output_weights.T) * a_h * (1 - a_h)
-    
-    # gradients for the output layer
-    output_weights_gradient = np.matmul(a_h.T, error_output)
-    output_bias_gradient = np.sum(error_output, axis=0)
-    
-    # gradient for the hidden layer
-    hidden_weights_gradient = np.matmul(X.T, error_hidden)
-    hidden_bias_gradient = np.sum(error_hidden, axis=0)
-
-    return output_weights_gradient, output_bias_gradient, hidden_weights_gradient, hidden_bias_gradient
-
-print("Old accuracy on training data: " + str(accuracy_score(predict(X_train), Y_train)))
-
-eta = 0.01
-lmbd = 0.01
-for i in range(1000):
-    # calculate gradients
-    dWo, dBo, dWh, dBh = backpropagation(X_train, Y_train_onehot)
-    
-    # regularization term gradients
-    dWo += lmbd * output_weights
-    dWh += lmbd * hidden_weights
-    
-    # update weights and biases
-    output_weights -= eta * dWo
-    output_bias -= eta * dBo
-    hidden_weights -= eta * dWh
-    hidden_bias -= eta * dBh
-
-print("New accuracy on training data: " + str(accuracy_score(predict(X_train), Y_train)))
-
    
 
    
 
    
 
@@ -393,7 +313,7 @@ lmbd = 0.0153
   
  • 54
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs045.html b/doc/pub/NeuralNet/html/._NeuralNet-bs045.html
index 6e6490f28..1331dd56d 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs045.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs045.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,23 +257,117 @@ MathJax.Hub.Config({
 
 
 
-
    Improving performance 
    
+
    Matrix  multiplication 
    
 
 
    
-As we can see the network does not seem to be learning at all. It seems to be just guessing the label for each image.  
-In order to obtain a network that does something useful, we will have to do a bit more work.
+To more efficently train our network these equations are implemented using matrix operations.  
+The error in the output layer is calculated simply as, with \( \hat{t} \) being our targets,  
+
+$$ \delta_L = \hat{t} - \hat{y} = (n_{inputs}, n_{categories}) .$$
 
 
    
-The choice of hyperparameters such as learning rate and regularization parameter is hugely influential for the performance of the network. Typically a grid-search is performed, wherein we test different hyperparameters separated by orders of magnitude. For example we could test the learning rates \( \eta = 10^{-6}, 10^{-5},...,10^{-1} \) with different regularization parameters \( \lambda = 10^{-6},...,10^{-0} \).
+The gradient for the output weights is calculated as  
+
+$$ \nabla W_{L} = \hat{a}^T \delta_L   = (n_{hidden}, n_{categories}) ,$$
 
 
    
-Next, we haven't implemented minibatching yet, which introduces stochasticity and is though to act as an important regularizer on the weights. We call a feed-forward + backward pass with a minibatch an iteration, and a full training period
-going through the entire dataset (\( n/M \) batches) an epoch.
+where \( \hat{a} = (n_{inputs}, n_{hidden}) \). This simply means that we are summing up the gradients for each input.  
+Since we are going backwards we have to transpose the activation matrix.
 
 
    
-If this does not improve network performance, you may want to consider altering the network architecture, adding more neurons or hidden layers.  
-Andrew Ng goes through some of these considerations in this video. You can find a summary of the video here.
+The gradient with respect to the output bias is then  
 
+$$ \nabla \hat{b}_{L} = \sum_{i=1}^{n_{inputs}} \delta_L = (n_{categories}) .$$
+
+
    
+The error in the hidden layer is  
+
+$$ \Delta_h = \delta_L W_{L}^T \circ f'(z_{h}) = \delta_L W_{L}^T \circ a_{h} \circ (1 - a_{h}) = (n_{inputs}, n_{hidden}) ,$$
+
+
    
+where \( f'(a_{h}) \) is the derivative of the activation in the hidden layer. The matrix products mean
+that we are summing up the products for each neuron in the output layer. The symbol \( \circ \) denotes
+the Hadamard product, meaning element-wise multiplication.
+
+
    
+This again gives us the gradients in the hidden layer:  
+
+$$ \nabla W_{h} = X^T \delta_h = (n_{features}, n_{hidden}) ,$$  
+
+$$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
+
+
    
+
+
+
    # to categorical turns our integer vector into a onehot representation
+from sklearn.metrics import accuracy_score
+
+# one-hot in numpy
+def to_categorical_numpy(integer_vector):
+    n_inputs = len(integer_vector)
+    n_categories = np.max(integer_vector) + 1
+    onehot_vector = np.zeros((n_inputs, n_categories))
+    onehot_vector[range(n_inputs), integer_vector] = 1
+    
+    return onehot_vector
+
+#Y_train_onehot, Y_test_onehot = to_categorical(Y_train), to_categorical(Y_test)
+Y_train_onehot, Y_test_onehot = to_categorical_numpy(Y_train), to_categorical_numpy(Y_test)
+
+def feed_forward_train(X):
+    # weighted sum of inputs to the hidden layer
+    z_h = np.matmul(X, hidden_weights) + hidden_bias
+    # activation in the hidden layer
+    a_h = sigmoid(z_h)
+    
+    # weighted sum of inputs to the output layer
+    z_o = np.matmul(a_h, output_weights) + output_bias
+    # softmax output
+    # axis 0 holds each input and axis 1 the probabilities of each category
+    exp_term = np.exp(z_o)
+    probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
+    
+    # for backpropagation need activations in hidden and output layers
+    return a_h, probabilities
+
+def backpropagation(X, Y):
+    a_h, probabilities = feed_forward_train(X)
+    
+    # error in the output layer
+    error_output = probabilities - Y
+    # error in the hidden layer
+    error_hidden = np.matmul(error_output, output_weights.T) * a_h * (1 - a_h)
+    
+    # gradients for the output layer
+    output_weights_gradient = np.matmul(a_h.T, error_output)
+    output_bias_gradient = np.sum(error_output, axis=0)
+    
+    # gradient for the hidden layer
+    hidden_weights_gradient = np.matmul(X.T, error_hidden)
+    hidden_bias_gradient = np.sum(error_hidden, axis=0)
+
+    return output_weights_gradient, output_bias_gradient, hidden_weights_gradient, hidden_bias_gradient
+
+print("Old accuracy on training data: " + str(accuracy_score(predict(X_train), Y_train)))
+
+eta = 0.01
+lmbd = 0.01
+for i in range(1000):
+    # calculate gradients
+    dWo, dBo, dWh, dBh = backpropagation(X_train, Y_train_onehot)
+    
+    # regularization term gradients
+    dWo += lmbd * output_weights
+    dWh += lmbd * hidden_weights
+    
+    # update weights and biases
+    output_weights -= eta * dWo
+    output_bias -= eta * dBo
+    hidden_weights -= eta * dWh
+    hidden_bias -= eta * dBh
+
+print("New accuracy on training data: " + str(accuracy_score(predict(X_train), Y_train)))
+
    
 
    
 
    
 
@@ -296,7 +394,7 @@ Andrew Ng goes through some of these considerations in this 54
   
  • 55
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs046.html b/doc/pub/NeuralNet/html/._NeuralNet-bs046.html
index 1edef29c4..762f31a53 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs046.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs046.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,116 +257,23 @@ MathJax.Hub.Config({
 
 
 
-
    Full object-oriented implementation 
    
+
    Improving performance 
    
 
 
    
-It is very natural to think of the network as an object, with specific instances of the network
-being realizations of this object with different hyperparameters. An implementation using Python classes provides a clean structure and interface, and the full implementation of our neural network is given below.
+As we can see the network does not seem to be learning at all. It seems to be just guessing the label for each image.  
+In order to obtain a network that does something useful, we will have to do a bit more work.
 
 
    
+The choice of hyperparameters such as learning rate and regularization parameter is hugely influential for the performance of the network. Typically a grid-search is performed, wherein we test different hyperparameters separated by orders of magnitude. For example we could test the learning rates \( \eta = 10^{-6}, 10^{-5},...,10^{-1} \) with different regularization parameters \( \lambda = 10^{-6},...,10^{-0} \).
 
-
-
    class NeuralNetwork:
-    def __init__(
-        self,
-        X_data,
-        Y_data,
-        n_hidden_neurons=50,
-        n_categories=10,
-        epochs=10,
-        batch_size=100,
-        eta=0.1,
-        lmbd=0.0,
+
    
+Next, we haven't implemented minibatching yet, which introduces stochasticity and is though to act as an important regularizer on the weights. We call a feed-forward + backward pass with a minibatch an iteration, and a full training period
+going through the entire dataset (\( n/M \) batches) an epoch.
 
-    ):
-        self.X_data_full = X_data
-        self.Y_data_full = Y_data
+
    
+If this does not improve network performance, you may want to consider altering the network architecture, adding more neurons or hidden layers.  
+Andrew Ng goes through some of these considerations in this video. You can find a summary of the video here.
 
-        self.n_inputs = X_data.shape[0]
-        self.n_features = X_data.shape[1]
-        self.n_hidden_neurons = n_hidden_neurons
-        self.n_categories = n_categories
-
-        self.epochs = epochs
-        self.batch_size = batch_size
-        self.iterations = self.n_inputs // self.batch_size
-        self.eta = eta
-        self.lmbd = lmbd
-
-        self.create_biases_and_weights()
-
-    def create_biases_and_weights(self):
-        self.hidden_weights = np.random.randn(self.n_features, self.n_hidden_neurons)
-        self.hidden_bias = np.zeros(self.n_hidden_neurons) + 0.01
-
-        self.output_weights = np.random.randn(self.n_hidden_neurons, self.n_categories)
-        self.output_bias = np.zeros(self.n_categories) + 0.01
-
-    def feed_forward(self):
-        # feed-forward for training
-        self.z_h = np.matmul(self.X_data, self.hidden_weights) + self.hidden_bias
-        self.a_h = sigmoid(self.z_h)
-
-        self.z_o = np.matmul(self.a_h, self.output_weights) + self.output_bias
-
-        exp_term = np.exp(self.z_o)
-        self.probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
-
-    def feed_forward_out(self, X):
-        # feed-forward for output
-        z_h = np.matmul(X, self.hidden_weights) + self.hidden_bias
-        a_h = sigmoid(z_h)
-
-        z_o = np.matmul(a_h, self.output_weights) + self.output_bias
-        
-        exp_term = np.exp(z_o)
-        probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
-        return probabilities
-
-    def backpropagation(self):
-        error_output = self.probabilities - self.Y_data
-        error_hidden = np.matmul(error_output, self.output_weights.T) * self.a_h * (1 - self.a_h)
-
-        self.output_weights_gradient = np.matmul(self.a_h.T, error_output)
-        self.output_bias_gradient = np.sum(error_output, axis=0)
-
-        self.hidden_weights_gradient = np.matmul(self.X_data.T, error_hidden)
-        self.hidden_bias_gradient = np.sum(error_hidden, axis=0)
-
-        if self.lmbd > 0.0:
-            self.output_weights_gradient += self.lmbd * self.output_weights
-            self.hidden_weights_gradient += self.lmbd * self.hidden_weights
-
-        self.output_weights -= self.eta * self.output_weights_gradient
-        self.output_bias -= self.eta * self.output_bias_gradient
-        self.hidden_weights -= self.eta * self.hidden_weights_gradient
-        self.hidden_bias -= self.eta * self.hidden_bias_gradient
-
-    def predict(self, X):
-        probabilities = self.feed_forward_out(X)
-        return np.argmax(probabilities, axis=1)
-
-    def predict_probabilities(self, X):
-        probabilities = self.feed_forward_out(X)
-        return probabilities
-
-    def train(self):
-        data_indices = np.arange(self.n_inputs)
-
-        for i in range(self.epochs):
-            for j in range(self.iterations):
-                # pick datapoints with replacement
-                chosen_datapoints = np.random.choice(
-                    data_indices, size=self.batch_size, replace=False
-                )
-
-                # minibatch training data
-                self.X_data = self.X_data_full[chosen_datapoints]
-                self.Y_data = self.Y_data_full[chosen_datapoints]
-
-                self.feed_forward()
-                self.backpropagation()
-
    
 
    
 
    
 
@@ -389,7 +300,7 @@ being realizations of this object with different hyperparameters. An implementat
   
  • 55
    • 
   
  • 56
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs047.html b/doc/pub/NeuralNet/html/._NeuralNet-bs047.html
index d18667bac..90bba8b4c 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs047.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs047.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,37 +257,115 @@ MathJax.Hub.Config({
 
 
 
-
    Evaluate model performance on test data 
    
+
    Full object-oriented implementation 
    
 
 
    
-To measure the performance of our network we evaluate how well it does it data it has never seen before, i.e. the test data.  
-We measure the performance of the network using the accuracy score.  
-The accuracy is as you would expect just the number of images correctly labeled divided by the total number of images. A perfect classifier will have an accuracy score of \( 1 \).  
-
-$$ \text{Accuracy} = \frac{\sum_{i=1}^n I(\hat{y}_i = y_i)}{n} ,$$
-
-
    
-where \( I \) is the indicator function, \( 1 \) if \( \hat{y}_i = y_i \) and \( 0 \) otherwise.
+It is very natural to think of the network as an object, with specific instances of the network
+being realizations of this object with different hyperparameters. An implementation using Python classes provides a clean structure and interface, and the full implementation of our neural network is given below.
 
 
    
 
 
-
    epochs = 100
-batch_size = 100
+
    class NeuralNetwork:
+    def __init__(
+        self,
+        X_data,
+        Y_data,
+        n_hidden_neurons=50,
+        n_categories=10,
+        epochs=10,
+        batch_size=100,
+        eta=0.1,
+        lmbd=0.0,
 
-dnn = NeuralNetwork(X_train, Y_train_onehot, eta=eta, lmbd=lmbd, epochs=epochs, batch_size=batch_size,
-                    n_hidden_neurons=n_hidden_neurons, n_categories=n_categories)
-dnn.train()
-test_predict = dnn.predict(X_test)
+    ):
+        self.X_data_full = X_data
+        self.Y_data_full = Y_data
 
-# accuracy score from scikit library
-print("Accuracy score on test set: ", accuracy_score(Y_test, test_predict))
+        self.n_inputs = X_data.shape[0]
+        self.n_features = X_data.shape[1]
+        self.n_hidden_neurons = n_hidden_neurons
+        self.n_categories = n_categories
 
-# equivalent in numpy
-def accuracy_score_numpy(Y_test, Y_pred):
-    return np.sum(Y_test == Y_pred) / len(Y_test)
+        self.epochs = epochs
+        self.batch_size = batch_size
+        self.iterations = self.n_inputs // self.batch_size
+        self.eta = eta
+        self.lmbd = lmbd
 
-#print("Accuracy score on test set: ", accuracy_score_numpy(Y_test, test_predict))
+        self.create_biases_and_weights()
+
+    def create_biases_and_weights(self):
+        self.hidden_weights = np.random.randn(self.n_features, self.n_hidden_neurons)
+        self.hidden_bias = np.zeros(self.n_hidden_neurons) + 0.01
+
+        self.output_weights = np.random.randn(self.n_hidden_neurons, self.n_categories)
+        self.output_bias = np.zeros(self.n_categories) + 0.01
+
+    def feed_forward(self):
+        # feed-forward for training
+        self.z_h = np.matmul(self.X_data, self.hidden_weights) + self.hidden_bias
+        self.a_h = sigmoid(self.z_h)
+
+        self.z_o = np.matmul(self.a_h, self.output_weights) + self.output_bias
+
+        exp_term = np.exp(self.z_o)
+        self.probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
+
+    def feed_forward_out(self, X):
+        # feed-forward for output
+        z_h = np.matmul(X, self.hidden_weights) + self.hidden_bias
+        a_h = sigmoid(z_h)
+
+        z_o = np.matmul(a_h, self.output_weights) + self.output_bias
+        
+        exp_term = np.exp(z_o)
+        probabilities = exp_term / np.sum(exp_term, axis=1, keepdims=True)
+        return probabilities
+
+    def backpropagation(self):
+        error_output = self.probabilities - self.Y_data
+        error_hidden = np.matmul(error_output, self.output_weights.T) * self.a_h * (1 - self.a_h)
+
+        self.output_weights_gradient = np.matmul(self.a_h.T, error_output)
+        self.output_bias_gradient = np.sum(error_output, axis=0)
+
+        self.hidden_weights_gradient = np.matmul(self.X_data.T, error_hidden)
+        self.hidden_bias_gradient = np.sum(error_hidden, axis=0)
+
+        if self.lmbd > 0.0:
+            self.output_weights_gradient += self.lmbd * self.output_weights
+            self.hidden_weights_gradient += self.lmbd * self.hidden_weights
+
+        self.output_weights -= self.eta * self.output_weights_gradient
+        self.output_bias -= self.eta * self.output_bias_gradient
+        self.hidden_weights -= self.eta * self.hidden_weights_gradient
+        self.hidden_bias -= self.eta * self.hidden_bias_gradient
+
+    def predict(self, X):
+        probabilities = self.feed_forward_out(X)
+        return np.argmax(probabilities, axis=1)
+
+    def predict_probabilities(self, X):
+        probabilities = self.feed_forward_out(X)
+        return probabilities
+
+    def train(self):
+        data_indices = np.arange(self.n_inputs)
+
+        for i in range(self.epochs):
+            for j in range(self.iterations):
+                # pick datapoints with replacement
+                chosen_datapoints = np.random.choice(
+                    data_indices, size=self.batch_size, replace=False
+                )
+
+                # minibatch training data
+                self.X_data = self.X_data_full[chosen_datapoints]
+                self.Y_data = self.Y_data_full[chosen_datapoints]
+
+                self.feed_forward()
+                self.backpropagation()
 
    
 
    
 
    
@@ -311,7 +393,7 @@ test_predict = dnn56
   
  • 57
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs048.html b/doc/pub/NeuralNet/html/._NeuralNet-bs048.html
index 5f716c042..77963059e 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs048.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs048.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,35 +257,37 @@ MathJax.Hub.Config({
 
 
 
-
    Adjust hyperparameters 
    
+
    Evaluate model performance on test data 
    
 
 
    
-We now perform a grid search to find the optimal hyperparameters for the network.  
-Note that we are only using 1 layer with 50 neurons, and human performance is estimated to be around \( 98\% \) (\( 2\% \) error rate).
+To measure the performance of our network we evaluate how well it does it data it has never seen before, i.e. the test data.  
+We measure the performance of the network using the accuracy score.  
+The accuracy is as you would expect just the number of images correctly labeled divided by the total number of images. A perfect classifier will have an accuracy score of \( 1 \).  
+
+$$ \text{Accuracy} = \frac{\sum_{i=1}^n I(\hat{y}_i = y_i)}{n} ,$$
+
+
    
+where \( I \) is the indicator function, \( 1 \) if \( \hat{y}_i = y_i \) and \( 0 \) otherwise.
 
 
    
 
 
-
    eta_vals = np.logspace(-5, 1, 7)
-lmbd_vals = np.logspace(-5, 1, 7)
-# store the models for later use
-DNN_numpy = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+
    epochs = 100
+batch_size = 100
 
-# grid search
-for i, eta in enumerate(eta_vals):
-    for j, lmbd in enumerate(lmbd_vals):
-        dnn = NeuralNetwork(X_train, Y_train_onehot, eta=eta, lmbd=lmbd, epochs=epochs, batch_size=batch_size,
-                            n_hidden_neurons=n_hidden_neurons, n_categories=n_categories)
-        dnn.train()
-        
-        DNN_numpy[i][j] = dnn
-        
-        test_predict = dnn.predict(X_test)
-        
-        print("Learning rate  = ", eta)
-        print("Lambda = ", lmbd)
-        print("Accuracy score on test set: ", accuracy_score(Y_test, test_predict))
-        print()
+dnn = NeuralNetwork(X_train, Y_train_onehot, eta=eta, lmbd=lmbd, epochs=epochs, batch_size=batch_size,
+                    n_hidden_neurons=n_hidden_neurons, n_categories=n_categories)
+dnn.train()
+test_predict = dnn.predict(X_test)
+
+# accuracy score from scikit library
+print("Accuracy score on test set: ", accuracy_score(Y_test, test_predict))
+
+# equivalent in numpy
+def accuracy_score_numpy(Y_test, Y_pred):
+    return np.sum(Y_test == Y_pred) / len(Y_test)
+
+#print("Accuracy score on test set: ", accuracy_score_numpy(Y_test, test_predict))
 
    
 
    
 
    
@@ -309,7 +315,7 @@ DNN_numpy = np.
   
  • 57
    • 
   
  • 58
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs049.html b/doc/pub/NeuralNet/html/._NeuralNet-bs049.html
index ad242d3bb..3815d10a1 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs049.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs049.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,44 +257,35 @@ MathJax.Hub.Config({
 
 
 
-
    Visualization 
    
+
    Adjust hyperparameters 
    
+
+
    
+We now perform a grid search to find the optimal hyperparameters for the network.  
+Note that we are only using 1 layer with 50 neurons, and human performance is estimated to be around \( 98\% \) (\( 2\% \) error rate).
 
 
    
 
 
-
    # visual representation of grid search
-# uses seaborn heatmap, you can also do this with matplotlib imshow
-import seaborn as sns
+
    eta_vals = np.logspace(-5, 1, 7)
+lmbd_vals = np.logspace(-5, 1, 7)
+# store the models for later use
+DNN_numpy = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
 
-sns.set()
-
-train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-
-for i in range(len(eta_vals)):
-    for j in range(len(lmbd_vals)):
-        dnn = DNN_numpy[i][j]
+# grid search
+for i, eta in enumerate(eta_vals):
+    for j, lmbd in enumerate(lmbd_vals):
+        dnn = NeuralNetwork(X_train, Y_train_onehot, eta=eta, lmbd=lmbd, epochs=epochs, batch_size=batch_size,
+                            n_hidden_neurons=n_hidden_neurons, n_categories=n_categories)
+        dnn.train()
         
-        train_pred = dnn.predict(X_train) 
-        test_pred = dnn.predict(X_test)
-
-        train_accuracy[i][j] = accuracy_score(Y_train, train_pred)
-        test_accuracy[i][j] = accuracy_score(Y_test, test_pred)
-
+        DNN_numpy[i][j] = dnn
         
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Training Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
-
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Test Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
+        test_predict = dnn.predict(X_test)
+        
+        print("Learning rate  = ", eta)
+        print("Lambda = ", lmbd)
+        print("Accuracy score on test set: ", accuracy_score(Y_test, test_predict))
+        print()
 
    
 
    
 
    
@@ -318,7 +313,7 @@ plt.show()
   
  • 58
    • 
   
  • 59
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs050.html b/doc/pub/NeuralNet/html/._NeuralNet-bs050.html
index b7c8ff88d..47b18cf03 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs050.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs050.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,41 +257,44 @@ MathJax.Hub.Config({
 
 
 
-
    scikit-learn implementation 
    
-
-
    
-scikit-learn focuses more
-on traditional machine learning methods, such as regression,
-clustering, decision trees, etc. As such, it has only two types of
-neural networks: Multi Layer Perceptron outputting continuous values,
-MPLRegressor, and Multi Layer Perceptron outputting labels,
-MLPClassifier. We will see how simple it is to use these classes.
-
-
    
-scikit-learn implements a few improvements from our neural network,
-such as early stopping, a varying learning rate, different
-optimization methods, etc. We would therefore expect a better
-performance overall.
+
    Visualization 
    
 
 
    
 
 
-
    from sklearn.neural_network import MLPClassifier
-# store models for later use
-DNN_scikit = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+
    # visual representation of grid search
+# uses seaborn heatmap, you can also do this with matplotlib imshow
+import seaborn as sns
 
-for i, eta in enumerate(eta_vals):
-    for j, lmbd in enumerate(lmbd_vals):
-        dnn = MLPClassifier(hidden_layer_sizes=(n_hidden_neurons), activation='logistic',
-                            alpha=lmbd, learning_rate_init=eta, max_iter=epochs)
-        dnn.fit(X_train, Y_train)
+sns.set()
+
+train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+
+for i in range(len(eta_vals)):
+    for j in range(len(lmbd_vals)):
+        dnn = DNN_numpy[i][j]
         
-        DNN_scikit[i][j] = dnn
+        train_pred = dnn.predict(X_train) 
+        test_pred = dnn.predict(X_test)
+
+        train_accuracy[i][j] = accuracy_score(Y_train, train_pred)
+        test_accuracy[i][j] = accuracy_score(Y_test, test_pred)
+
         
-        print("Learning rate  = ", eta)
-        print("Lambda = ", lmbd)
-        print("Accuracy score on test set: ", dnn.score(X_test, Y_test))
-        print()
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Training Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
+
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Test Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
 
    
 
    
 
    
@@ -315,7 +322,7 @@ DNN_scikit = np
   
  • 59
    • 
   
  • 60
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs051.html b/doc/pub/NeuralNet/html/._NeuralNet-bs051.html
index 2cc0b8b79..669324e88 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs051.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs051.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,44 +257,41 @@ MathJax.Hub.Config({
 
 
 
-
    Visualization 
    
+
    scikit-learn implementation 
    
+
+
    
+scikit-learn focuses more
+on traditional machine learning methods, such as regression,
+clustering, decision trees, etc. As such, it has only two types of
+neural networks: Multi Layer Perceptron outputting continuous values,
+MPLRegressor, and Multi Layer Perceptron outputting labels,
+MLPClassifier. We will see how simple it is to use these classes.
+
+
    
+scikit-learn implements a few improvements from our neural network,
+such as early stopping, a varying learning rate, different
+optimization methods, etc. We would therefore expect a better
+performance overall.
+
 
    
 
 
-
    # optional
-# visual representation of grid search
-# uses seaborn heatmap, could probably do this in matplotlib
-import seaborn as sns
+
    from sklearn.neural_network import MLPClassifier
+# store models for later use
+DNN_scikit = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
 
-sns.set()
-
-train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-
-for i in range(len(eta_vals)):
-    for j in range(len(lmbd_vals)):
-        dnn = DNN_scikit[i][j]
+for i, eta in enumerate(eta_vals):
+    for j, lmbd in enumerate(lmbd_vals):
+        dnn = MLPClassifier(hidden_layer_sizes=(n_hidden_neurons), activation='logistic',
+                            alpha=lmbd, learning_rate_init=eta, max_iter=epochs)
+        dnn.fit(X_train, Y_train)
         
-        train_pred = dnn.predict(X_train) 
-        test_pred = dnn.predict(X_test)
-
-        train_accuracy[i][j] = accuracy_score(Y_train, train_pred)
-        test_accuracy[i][j] = accuracy_score(Y_test, test_pred)
-
+        DNN_scikit[i][j] = dnn
         
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Training Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
-
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Test Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
+        print("Learning rate  = ", eta)
+        print("Lambda = ", lmbd)
+        print("Accuracy score on test set: ", dnn.score(X_test, Y_test))
+        print()
 
    
 
    
 
    
@@ -318,7 +319,7 @@ plt.show()
   
  • 60
    • 
   
  • 61
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs052.html b/doc/pub/NeuralNet/html/._NeuralNet-bs052.html
index 258d88c7e..bddb22bf1 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs052.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs052.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,18 +257,45 @@ MathJax.Hub.Config({
 
 
 
-
    Building neural networks in Tensorflow and Keras 
    
-
+
    Visualization 
    
 
    
-Now we want  to build on the experience gained from our neural network implementation in NumPy and scikit-learn
-and use it to construct a neural network in Tensorflow. Once we have constructed a neural network in NumPy
-and Tensorflow, building one in Keras is really quite trivial, though the performance may suffer.
 
-
    
-In our previous example we used only one hidden layer, and in this we will use two. From this it should be quite
-clear how to build one using an arbitrary number of hidden layers, using data structures such as Python lists or
-NumPy arrays.
+
+
    # optional
+# visual representation of grid search
+# uses seaborn heatmap, could probably do this in matplotlib
+import seaborn as sns
 
+sns.set()
+
+train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+
+for i in range(len(eta_vals)):
+    for j in range(len(lmbd_vals)):
+        dnn = DNN_scikit[i][j]
+        
+        train_pred = dnn.predict(X_train) 
+        test_pred = dnn.predict(X_test)
+
+        train_accuracy[i][j] = accuracy_score(Y_train, train_pred)
+        test_accuracy[i][j] = accuracy_score(Y_test, test_pred)
+
+        
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Training Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
+
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Test Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
+
    
 
    
 
    
 
@@ -291,7 +322,7 @@ NumPy arrays.
   
  • 61
    • 
   
  • 62
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs053.html b/doc/pub/NeuralNet/html/._NeuralNet-bs053.html
index b7a2ad902..063cc6412 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs053.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs053.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,48 +257,18 @@ MathJax.Hub.Config({
 
 
 
-
    Tensorflow 
    
+
    Building neural networks in Tensorflow and Keras 
    
 
 
    
-Tensorflow is an open source library machine learning library
-developed by the Google Brain team for internal use. It was released
-under the Apache 2.0 open source license in November 9, 2015.
+Now we want  to build on the experience gained from our neural network implementation in NumPy and scikit-learn
+and use it to construct a neural network in Tensorflow. Once we have constructed a neural network in NumPy
+and Tensorflow, building one in Keras is really quite trivial, though the performance may suffer.
 
 
    
-Tensorflow is a computational framework that allows you to construct
-machine learning models at different levels of abstraction, from
-high-level, object-oriented APIs like Keras, down to the C++ kernels
-that Tensorflow is built upon. The higher levels of abstraction are
-simpler to use, but less flexible, and our choice of implementation
-should reflect the problems we are trying to solve.
+In our previous example we used only one hidden layer, and in this we will use two. From this it should be quite
+clear how to build one using an arbitrary number of hidden layers, using data structures such as Python lists or
+NumPy arrays.
 
-
    
-Tensorflow uses so-called graphs to represent your computation
-in terms of the dependencies between individual operations, such that you first build a Tensorflow graph
-to represent your model, and then create a Tensorflow session to run the graph.
-
-
    
-In this guide we will analyze the same data as we did in our NumPy and
-scikit-learn tutorial, gathered from the MNIST database of images. We
-will give an introduction to the lower level Python Application
-Program Interfaces (APIs), and see how we use them to build our graph.
-Then we will build (effectively) the same graph in Keras, to see just
-how simple solving a machine learning problem can be.
-
-
    
-To install tensorflow on Unix/Linux systems, use pip as
-
    
-
-
-
    pip3 install tensorflow
-
    
-
    
-and/or if you use anaconda, just write (or install from the graphical user interface)
-
    
-
-
-
    conda install tensorflow
-
    
 
    
 
    
 
@@ -321,7 +295,7 @@ and/or if you use anaconda, just write (or install from the graphical use
   
  • 62
    • 
   
  • 63
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs054.html b/doc/pub/NeuralNet/html/._NeuralNet-bs054.html
index 8eb8c1e41..3a00235c3 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs054.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs054.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,68 +257,47 @@ MathJax.Hub.Config({
 
 
 
-
    Collect and pre-process data 
    
+
    Tensorflow 
    
 
+
    
+Tensorflow is an open source library machine learning library
+developed by the Google Brain team for internal use. It was released
+under the Apache 2.0 open source license in November 9, 2015.
+
+
    
+Tensorflow is a computational framework that allows you to construct
+machine learning models at different levels of abstraction, from
+high-level, object-oriented APIs like Keras, down to the C++ kernels
+that Tensorflow is built upon. The higher levels of abstraction are
+simpler to use, but less flexible, and our choice of implementation
+should reflect the problems we are trying to solve.
+
+
    
+Tensorflow uses so-called graphs to represent your computation
+in terms of the dependencies between individual operations, such that you first build a Tensorflow graph
+to represent your model, and then create a Tensorflow session to run the graph.
+
+
    
+In this guide we will analyze the same data as we did in our NumPy and
+scikit-learn tutorial, gathered from the MNIST database of images. We
+will give an introduction to the lower level Python Application
+Program Interfaces (APIs), and see how we use them to build our graph.
+Then we will build (effectively) the same graph in Keras, to see just
+how simple solving a machine learning problem can be.
+
+
    
+To install tensorflow on Unix/Linux systems, use pip as
 
    
 
 
-
    # import necessary packages
-import numpy as np
-import matplotlib.pyplot as plt
-from sklearn import datasets
-
-
-# ensure the same random numbers appear every time
-np.random.seed(0)
-
-# display images in notebook
-%matplotlib inline
-plt.rcParams['figure.figsize'] = (12,12)
-
-
-# download MNIST dataset
-digits = datasets.load_digits()
-
-# define inputs and labels
-inputs = digits.images
-labels = digits.target
-
-print("inputs = (n_inputs, pixel_width, pixel_height) = " + str(inputs.shape))
-print("labels = (n_inputs) = " + str(labels.shape))
-
-
-# flatten the image
-# the value -1 means dimension is inferred from the remaining dimensions: 8x8 = 64
-n_inputs = len(inputs)
-inputs = inputs.reshape(n_inputs, -1)
-print("X = (n_inputs, n_features) = " + str(inputs.shape))
-
-
-# choose some random images to display
-indices = np.arange(n_inputs)
-random_indices = np.random.choice(indices, size=5)
-
-for i, image in enumerate(digits.images[random_indices]):
-    plt.subplot(1, 5, i+1)
-    plt.axis('off')
-    plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
-    plt.title("Label: %d" % digits.target[random_indices[i]])
-plt.show()
+
    pip3 install tensorflow
 
    
 
    
+and/or if you use anaconda, just write (or install from the graphical user interface)
+
    
 
 
-
    from keras.utils import to_categorical
-from sklearn.model_selection import train_test_split
-
-# one-hot representation of labels
-labels = to_categorical(labels)
-
-# split into train and test data
-train_size = 0.8
-test_size = 1 - train_size
-X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=train_size,
-                                                    test_size=test_size)
+
    conda install tensorflow
 
    
 
    
 
    
@@ -341,6 +324,8 @@ X_train, X_test, Y_train, Y_test = train_tes
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • ...
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs055.html b/doc/pub/NeuralNet/html/._NeuralNet-bs055.html
index 4433ef781..d3b4f78e5 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs055.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs055.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,147 +257,68 @@ MathJax.Hub.Config({
 
 
 
-
    Using TensorFlow backend 
    
-
-
      
-
    1.  Define model and architecture
      2. 
-
    2.  Choose cost function and optimizer
      3. 
-
    
+
    Collect and pre-process data 
    
 
 
    
 
 
-
    import tensorflow as tf
+
    # import necessary packages
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn import datasets
 
-class NeuralNetworkTensorflow:
-    def __init__(
-        self,
-        X_train,
-        Y_train,
-        X_test,
-        Y_test,
-        n_neurons_layer1=100,
-        n_neurons_layer2=50,
-        n_categories=2,
-        epochs=10,
-        batch_size=100,
-        eta=0.1,
-        lmbd=0.0,
-    ):
-        
-        # keep track of number of steps
-        self.global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
-        
-        self.X_train = X_train
-        self.Y_train = Y_train
-        self.X_test = X_test
-        self.Y_test = Y_test
-        
-        self.n_inputs = X_train.shape[0]
-        self.n_features = X_train.shape[1]
-        self.n_neurons_layer1 = n_neurons_layer1
-        self.n_neurons_layer2 = n_neurons_layer2
-        self.n_categories = n_categories
-        
-        self.epochs = epochs
-        self.batch_size = batch_size
-        self.iterations = self.n_inputs // self.batch_size
-        self.eta = eta
-        self.lmbd = lmbd
-        
-        # build network piece by piece
-        # name scopes (with) are used to enforce creation of new variables
-        # https://www.tensorflow.org/guide/variables
-        self.create_placeholders()
-        self.create_DNN()
-        self.create_loss()
-        self.create_optimiser()
-        self.create_accuracy()
-    
-    def create_placeholders(self):
-        # placeholders are fine here, but "Datasets" are the preferred method
-        # of streaming data into a model
-        with tf.name_scope('data'):
-            self.X = tf.placeholder(tf.float32, shape=(None, self.n_features), name='X_data')
-            self.Y = tf.placeholder(tf.float32, shape=(None, self.n_categories), name='Y_data')
-    
-    def create_DNN(self):
-        with tf.name_scope('DNN'):
-            # the weights are stored to calculate regularization loss later
-            
-            # Fully connected layer 1
-            self.W_fc1 = self.weight_variable([self.n_features, self.n_neurons_layer1], name='fc1', dtype=tf.float32)
-            b_fc1 = self.bias_variable([self.n_neurons_layer1], name='fc1', dtype=tf.float32)
-            a_fc1 = tf.nn.sigmoid(tf.matmul(self.X, self.W_fc1) + b_fc1)
-            
-            # Fully connected layer 2
-            self.W_fc2 = self.weight_variable([self.n_neurons_layer1, self.n_neurons_layer2], name='fc2', dtype=tf.float32)
-            b_fc2 = self.bias_variable([self.n_neurons_layer2], name='fc2', dtype=tf.float32)
-            a_fc2 = tf.nn.sigmoid(tf.matmul(a_fc1, self.W_fc2) + b_fc2)
-            
-            # Output layer
-            self.W_out = self.weight_variable([self.n_neurons_layer2, self.n_categories], name='out', dtype=tf.float32)
-            b_out = self.bias_variable([self.n_categories], name='out', dtype=tf.float32)
-            self.z_out = tf.matmul(a_fc2, self.W_out) + b_out
-    
-    def create_loss(self):
-        with tf.name_scope('loss'):
-            softmax_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.Y, logits=self.z_out))
-            
-            regularizer_loss_fc1 = tf.nn.l2_loss(self.W_fc1)
-            regularizer_loss_fc2 = tf.nn.l2_loss(self.W_fc2)
-            regularizer_loss_out = tf.nn.l2_loss(self.W_out)
-            regularizer_loss = self.lmbd*(regularizer_loss_fc1 + regularizer_loss_fc2 + regularizer_loss_out)
-            
-            self.loss = softmax_loss + regularizer_loss
 
-    def create_accuracy(self):
-        with tf.name_scope('accuracy'):
-            probabilities = tf.nn.softmax(self.z_out)
-            predictions = tf.argmax(probabilities, axis=1)
-            labels = tf.argmax(self.Y, axis=1)
-            
-            correct_predictions = tf.equal(predictions, labels)
-            correct_predictions = tf.cast(correct_predictions, tf.float32)
-            self.accuracy = tf.reduce_mean(correct_predictions)
-    
-    def create_optimiser(self):
-        with tf.name_scope('optimizer'):
-            self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.eta).minimize(self.loss, global_step=self.global_step)
-            
-    def weight_variable(self, shape, name='', dtype=tf.float32):
-        initial = tf.truncated_normal(shape, stddev=0.1)
-        return tf.Variable(initial, name=name, dtype=dtype)
-    
-    def bias_variable(self, shape, name='', dtype=tf.float32):
-        initial = tf.constant(0.1, shape=shape)
-        return tf.Variable(initial, name=name, dtype=dtype)
-    
-    def fit(self):
-        data_indices = np.arange(self.n_inputs)
+# ensure the same random numbers appear every time
+np.random.seed(0)
 
-        with tf.Session() as sess:
-            sess.run(tf.global_variables_initializer())
-            for i in range(self.epochs):
-                for j in range(self.iterations):
-                    chosen_datapoints = np.random.choice(data_indices, size=self.batch_size, replace=False)
-                    batch_X, batch_Y = self.X_train[chosen_datapoints], self.Y_train[chosen_datapoints]
-            
-                    sess.run([DNN.loss, DNN.optimizer],
-                        feed_dict={DNN.X: batch_X,
-                                   DNN.Y: batch_Y})
-                    accuracy = sess.run(DNN.accuracy,
-                        feed_dict={DNN.X: batch_X,
-                                   DNN.Y: batch_Y})
-                    step = sess.run(DNN.global_step)
-    
-            self.train_loss, self.train_accuracy = sess.run([DNN.loss, DNN.accuracy],
-                feed_dict={DNN.X: self.X_train,
-                           DNN.Y: self.Y_train})
-        
-            self.test_loss, self.test_accuracy = sess.run([DNN.loss, DNN.accuracy],
-                feed_dict={DNN.X: self.X_test,
-                           DNN.Y: self.Y_test})
+# display images in notebook
+%matplotlib inline
+plt.rcParams['figure.figsize'] = (12,12)
+
+
+# download MNIST dataset
+digits = datasets.load_digits()
+
+# define inputs and labels
+inputs = digits.images
+labels = digits.target
+
+print("inputs = (n_inputs, pixel_width, pixel_height) = " + str(inputs.shape))
+print("labels = (n_inputs) = " + str(labels.shape))
+
+
+# flatten the image
+# the value -1 means dimension is inferred from the remaining dimensions: 8x8 = 64
+n_inputs = len(inputs)
+inputs = inputs.reshape(n_inputs, -1)
+print("X = (n_inputs, n_features) = " + str(inputs.shape))
+
+
+# choose some random images to display
+indices = np.arange(n_inputs)
+random_indices = np.random.choice(indices, size=5)
+
+for i, image in enumerate(digits.images[random_indices]):
+    plt.subplot(1, 5, i+1)
+    plt.axis('off')
+    plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
+    plt.title("Label: %d" % digits.target[random_indices[i]])
+plt.show()
+
    
+
    
+
+
+
    from keras.utils import to_categorical
+from sklearn.model_selection import train_test_split
+
+# one-hot representation of labels
+labels = to_categorical(labels)
+
+# split into train and test data
+train_size = 0.8
+test_size = 1 - train_size
+X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=train_size,
+                                                    test_size=test_size)
 
    
 
    
 
    
@@ -419,6 +344,9 @@ MathJax.Hub.Config({
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • ...
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs056.html b/doc/pub/NeuralNet/html/._NeuralNet-bs056.html
index 5bd2150b5..d5471dc7e 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs056.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs056.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,80 +257,147 @@ MathJax.Hub.Config({
 
 
 
-
    Optimizing and using gradient descent 
    
+
    Using TensorFlow backend 
    
+
+
      
+
    1.  Define model and architecture
      2. 
+
    2.  Choose cost function and optimizer
      3. 
+
    
 
 
    
 
 
-
    epochs = 100
-batch_size = 100
-n_neurons_layer1 = 100
-n_neurons_layer2 = 50
-n_categories = 10
-eta_vals = np.logspace(-5, 1, 7)
-lmbd_vals = np.logspace(-5, 1, 7)
-
    
-
    
+
    import tensorflow as tf
 
-
-
    DNN_tf = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+class NeuralNetworkTensorflow:
+    def __init__(
+        self,
+        X_train,
+        Y_train,
+        X_test,
+        Y_test,
+        n_neurons_layer1=100,
+        n_neurons_layer2=50,
+        n_categories=2,
+        epochs=10,
+        batch_size=100,
+        eta=0.1,
+        lmbd=0.0,
+    ):
         
-for i, eta in enumerate(eta_vals):
-    for j, lmbd in enumerate(lmbd_vals):
-        DNN = NeuralNetworkTensorflow(X_train, Y_train, X_test, Y_test,
-                                      n_neurons_layer1, n_neurons_layer2, n_categories,
-                                      epochs=epochs, batch_size=batch_size, eta=eta, lmbd=lmbd)
-        DNN.fit()
+        # keep track of number of steps
+        self.global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
         
-        DNN_tf[i][j] = DNN
+        self.X_train = X_train
+        self.Y_train = Y_train
+        self.X_test = X_test
+        self.Y_test = Y_test
         
-        print("Learning rate = ", eta)
-        print("Lambda = ", lmbd)
-        print("Test accuracy: %.3f" % DNN.test_accuracy)
-        print()
-
    
-
    
-
-
-
    # optional
-# visual representation of grid search
-# uses seaborn heatmap, could probably do this in matplotlib
-import seaborn as sns
-
-sns.set()
-
-train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
-
-for i in range(len(eta_vals)):
-    for j in range(len(lmbd_vals)):
-        DNN = DNN_tf[i][j]
-
-        train_accuracy[i][j] = DNN.train_accuracy
-        test_accuracy[i][j] = DNN.test_accuracy
-
+        self.n_inputs = X_train.shape[0]
+        self.n_features = X_train.shape[1]
+        self.n_neurons_layer1 = n_neurons_layer1
+        self.n_neurons_layer2 = n_neurons_layer2
+        self.n_categories = n_categories
         
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Training Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
+        self.epochs = epochs
+        self.batch_size = batch_size
+        self.iterations = self.n_inputs // self.batch_size
+        self.eta = eta
+        self.lmbd = lmbd
+        
+        # build network piece by piece
+        # name scopes (with) are used to enforce creation of new variables
+        # https://www.tensorflow.org/guide/variables
+        self.create_placeholders()
+        self.create_DNN()
+        self.create_loss()
+        self.create_optimiser()
+        self.create_accuracy()
+    
+    def create_placeholders(self):
+        # placeholders are fine here, but "Datasets" are the preferred method
+        # of streaming data into a model
+        with tf.name_scope('data'):
+            self.X = tf.placeholder(tf.float32, shape=(None, self.n_features), name='X_data')
+            self.Y = tf.placeholder(tf.float32, shape=(None, self.n_categories), name='Y_data')
+    
+    def create_DNN(self):
+        with tf.name_scope('DNN'):
+            # the weights are stored to calculate regularization loss later
+            
+            # Fully connected layer 1
+            self.W_fc1 = self.weight_variable([self.n_features, self.n_neurons_layer1], name='fc1', dtype=tf.float32)
+            b_fc1 = self.bias_variable([self.n_neurons_layer1], name='fc1', dtype=tf.float32)
+            a_fc1 = tf.nn.sigmoid(tf.matmul(self.X, self.W_fc1) + b_fc1)
+            
+            # Fully connected layer 2
+            self.W_fc2 = self.weight_variable([self.n_neurons_layer1, self.n_neurons_layer2], name='fc2', dtype=tf.float32)
+            b_fc2 = self.bias_variable([self.n_neurons_layer2], name='fc2', dtype=tf.float32)
+            a_fc2 = tf.nn.sigmoid(tf.matmul(a_fc1, self.W_fc2) + b_fc2)
+            
+            # Output layer
+            self.W_out = self.weight_variable([self.n_neurons_layer2, self.n_categories], name='out', dtype=tf.float32)
+            b_out = self.bias_variable([self.n_categories], name='out', dtype=tf.float32)
+            self.z_out = tf.matmul(a_fc2, self.W_out) + b_out
+    
+    def create_loss(self):
+        with tf.name_scope('loss'):
+            softmax_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.Y, logits=self.z_out))
+            
+            regularizer_loss_fc1 = tf.nn.l2_loss(self.W_fc1)
+            regularizer_loss_fc2 = tf.nn.l2_loss(self.W_fc2)
+            regularizer_loss_out = tf.nn.l2_loss(self.W_out)
+            regularizer_loss = self.lmbd*(regularizer_loss_fc1 + regularizer_loss_fc2 + regularizer_loss_out)
+            
+            self.loss = softmax_loss + regularizer_loss
 
-fig, ax = plt.subplots(figsize = (10, 10))
-sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
-ax.set_title("Test Accuracy")
-ax.set_ylabel("$\eta$")
-ax.set_xlabel("$\lambda$")
-plt.show()
-
    
-
    
+    def create_accuracy(self):
+        with tf.name_scope('accuracy'):
+            probabilities = tf.nn.softmax(self.z_out)
+            predictions = tf.argmax(probabilities, axis=1)
+            labels = tf.argmax(self.Y, axis=1)
+            
+            correct_predictions = tf.equal(predictions, labels)
+            correct_predictions = tf.cast(correct_predictions, tf.float32)
+            self.accuracy = tf.reduce_mean(correct_predictions)
+    
+    def create_optimiser(self):
+        with tf.name_scope('optimizer'):
+            self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.eta).minimize(self.loss, global_step=self.global_step)
+            
+    def weight_variable(self, shape, name='', dtype=tf.float32):
+        initial = tf.truncated_normal(shape, stddev=0.1)
+        return tf.Variable(initial, name=name, dtype=dtype)
+    
+    def bias_variable(self, shape, name='', dtype=tf.float32):
+        initial = tf.constant(0.1, shape=shape)
+        return tf.Variable(initial, name=name, dtype=dtype)
+    
+    def fit(self):
+        data_indices = np.arange(self.n_inputs)
 
-
-
    # optional
-# we can use log files to visualize our graph in Tensorboard
-writer = tf.summary.FileWriter('logs/')
-writer.add_graph(tf.get_default_graph())
+        with tf.Session() as sess:
+            sess.run(tf.global_variables_initializer())
+            for i in range(self.epochs):
+                for j in range(self.iterations):
+                    chosen_datapoints = np.random.choice(data_indices, size=self.batch_size, replace=False)
+                    batch_X, batch_Y = self.X_train[chosen_datapoints], self.Y_train[chosen_datapoints]
+            
+                    sess.run([DNN.loss, DNN.optimizer],
+                        feed_dict={DNN.X: batch_X,
+                                   DNN.Y: batch_Y})
+                    accuracy = sess.run(DNN.accuracy,
+                        feed_dict={DNN.X: batch_X,
+                                   DNN.Y: batch_Y})
+                    step = sess.run(DNN.global_step)
+    
+            self.train_loss, self.train_accuracy = sess.run([DNN.loss, DNN.accuracy],
+                feed_dict={DNN.X: self.X_train,
+                           DNN.Y: self.Y_train})
+        
+            self.test_loss, self.test_accuracy = sess.run([DNN.loss, DNN.accuracy],
+                feed_dict={DNN.X: self.X_test,
+                           DNN.Y: self.Y_test})
 
    
 
    
 
    
@@ -351,6 +422,8 @@ writer.add_graph(tf62
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs057.html b/doc/pub/NeuralNet/html/._NeuralNet-bs057.html
index 3eac3d8b1..049c7039c 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs057.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs057.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,65 +257,36 @@ MathJax.Hub.Config({
 
 
 
-
    Using Keras 
    
-
-
    
-Keras is a high level neural network
-that supports Tensorflow, CTNK and Theano as backends.  
-If you have Tensorflow installed Keras is available through the tf.keras module.  
-If you have Anaconda installed you may run the following command
-
    
-
-
-
    conda install keras
-
    
-
    
-Alternatively, if you have Tensorflow or one of the other supported backends install you may use the pip package manager:
+
    Optimizing and using gradient descent 
    
 
 
    
 
 
-
    pip3 install keras
-
    
-
    
-or look up the instructions here.
-
-
    
-
-
-
    from keras.models import Sequential
-from keras.layers import Dense
-from keras.regularizers import l2
-from keras.optimizers import SGD
-
-def create_neural_network_keras(n_neurons_layer1, n_neurons_layer2, n_categories, eta, lmbd):
-    model = Sequential()
-    model.add(Dense(n_neurons_layer1, activation='sigmoid', kernel_regularizer=l2(lmbd)))
-    model.add(Dense(n_neurons_layer2, activation='sigmoid', kernel_regularizer=l2(lmbd)))
-    model.add(Dense(n_categories, activation='softmax'))
-    
-    sgd = SGD(lr=eta)
-    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
-    
-    return model
+
    epochs = 100
+batch_size = 100
+n_neurons_layer1 = 100
+n_neurons_layer2 = 50
+n_categories = 10
+eta_vals = np.logspace(-5, 1, 7)
+lmbd_vals = np.logspace(-5, 1, 7)
 
    
 
    
 
 
-
    DNN_keras = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+
    DNN_tf = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
         
 for i, eta in enumerate(eta_vals):
     for j, lmbd in enumerate(lmbd_vals):
-        DNN = create_neural_network_keras(n_neurons_layer1, n_neurons_layer2, n_categories,
-                                         eta=eta, lmbd=lmbd)
-        DNN.fit(X_train, Y_train, epochs=epochs, batch_size=batch_size, verbose=0)
-        scores = DNN.evaluate(X_test, Y_test)
+        DNN = NeuralNetworkTensorflow(X_train, Y_train, X_test, Y_test,
+                                      n_neurons_layer1, n_neurons_layer2, n_categories,
+                                      epochs=epochs, batch_size=batch_size, eta=eta, lmbd=lmbd)
+        DNN.fit()
         
-        DNN_keras[i][j] = DNN
+        DNN_tf[i][j] = DNN
         
         print("Learning rate = ", eta)
         print("Lambda = ", lmbd)
-        print("Test accuracy: %.3f" % scores[1])
+        print("Test accuracy: %.3f" % DNN.test_accuracy)
         print()
 
    
 
    
@@ -329,10 +304,10 @@ test_accuracy = npfor i in range(len(eta_vals)):
     for j in range(len(lmbd_vals)):
-        DNN = DNN_keras[i][j]
+        DNN = DNN_tf[i][j]
 
-        train_accuracy[i][j] = DNN.evaluate(X_train, Y_train)[1]
-        test_accuracy[i][j] = DNN.evaluate(X_test, Y_test)[1]
+        train_accuracy[i][j] = DNN.train_accuracy
+        test_accuracy[i][j] = DNN.test_accuracy
 
         
 fig, ax = plt.subplots(figsize = (10, 10))
@@ -350,6 +325,14 @@ ax.set_xlabel(&
 plt.show()
 
    
 
    
+
+
+
    # optional
+# we can use log files to visualize our graph in Tensorboard
+writer = tf.summary.FileWriter('logs/')
+writer.add_graph(tf.get_default_graph())
+
    
+
    
 
    
 
 
    
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs058.html b/doc/pub/NeuralNet/html/._NeuralNet-bs058.html
index dbe3a0dca..2da05d923 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs058.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs058.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,25 +255,104 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    Which activation function should I use? 
    
+
    Using Keras 
    
 
 
    
-Backpropagation algorithm works by going from the output layer to the
-input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the
-cost function with regards to each parameter in the network, it uses these gradients to update each
-parameter with a Gradient Descent step.
+Keras is a high level neural network
+that supports Tensorflow, CTNK and Theano as backends.  
+If you have Tensorflow installed Keras is available through the tf.keras module.  
+If you have Anaconda installed you may run the following command
+
    
+
+
+
    conda install keras
+
    
+
    
+Alternatively, if you have Tensorflow or one of the other supported backends install you may use the pip package manager:
 
 
    
-Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower
-layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually
-unchanged, and training never converges to a good solution. This is called the vanishing gradients
-problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many
-layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients
-problem, which is mostly encountered in recurrent neural networks. More generally,
-deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds
 
+
+
    pip3 install keras
+
    
+
    
+or look up the instructions here.
+
+
    
+
+
+
    from keras.models import Sequential
+from keras.layers import Dense
+from keras.regularizers import l2
+from keras.optimizers import SGD
+
+def create_neural_network_keras(n_neurons_layer1, n_neurons_layer2, n_categories, eta, lmbd):
+    model = Sequential()
+    model.add(Dense(n_neurons_layer1, activation='sigmoid', kernel_regularizer=l2(lmbd)))
+    model.add(Dense(n_neurons_layer2, activation='sigmoid', kernel_regularizer=l2(lmbd)))
+    model.add(Dense(n_categories, activation='softmax'))
+    
+    sgd = SGD(lr=eta)
+    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
+    
+    return model
+
    
+
    
+
+
+
    DNN_keras = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+        
+for i, eta in enumerate(eta_vals):
+    for j, lmbd in enumerate(lmbd_vals):
+        DNN = create_neural_network_keras(n_neurons_layer1, n_neurons_layer2, n_categories,
+                                         eta=eta, lmbd=lmbd)
+        DNN.fit(X_train, Y_train, epochs=epochs, batch_size=batch_size, verbose=0)
+        scores = DNN.evaluate(X_test, Y_test)
+        
+        DNN_keras[i][j] = DNN
+        
+        print("Learning rate = ", eta)
+        print("Lambda = ", lmbd)
+        print("Test accuracy: %.3f" % scores[1])
+        print()
+
    
+
    
+
+
+
    # optional
+# visual representation of grid search
+# uses seaborn heatmap, could probably do this in matplotlib
+import seaborn as sns
+
+sns.set()
+
+train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
+
+for i in range(len(eta_vals)):
+    for j in range(len(lmbd_vals)):
+        DNN = DNN_keras[i][j]
+
+        train_accuracy[i][j] = DNN.evaluate(X_train, Y_train)[1]
+        test_accuracy[i][j] = DNN.evaluate(X_test, Y_test)[1]
+
+        
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(train_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Training Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
+
+fig, ax = plt.subplots(figsize = (10, 10))
+sns.heatmap(test_accuracy, annot=True, ax=ax, cmap="viridis")
+ax.set_title("Test Accuracy")
+ax.set_ylabel("$\eta$")
+ax.set_xlabel("$\lambda$")
+plt.show()
+
    
 
    
 
    
 
@@ -291,6 +374,8 @@ deep neural networks suffer from unstable gradients, different layers may learn
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs059.html b/doc/pub/NeuralNet/html/._NeuralNet-bs059.html
index d6cd63cd1..13d0e4d5e 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs059.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs059.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,23 +257,31 @@ MathJax.Hub.Config({
 
 
 
-
    Is the Logistic activation function (Sigmoid)  our choice? 
    
+
    Which activation function should I use? 
    
 
 
    
-Although this unfortunate behavior has been empirically observed for quite a while (it was one of the
-reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that
-significant progress was made in understanding it.
+The Back propagation algorithm we derived above works by going from
+the output layer to the input layer, propagating the error gradient on
+the way. Once the algorithm has computed the gradient of the cost
+function with regards to each parameter in the network, it uses these
+gradients to update each parameter with a Gradient Descent (GD) step.
 
 
    
-A paper titled Understanding the Difficulty of Training Deep Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio1 found a few suspects,
-including the combination of the popular logistic sigmoid activation function and the weight initialization
-technique that was most popular at the time, namely random initialization using a normal distribution with
-a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this
-initialization scheme, the variance of the outputs of each layer is much greater than the variance of its
-inputs. Going forward in the network, the variance keeps increasing after each layer until the activation
-function saturates at the top layers. This is actually made worse by the fact that the logistic function has a
-mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the
-logistic function in deep networks).
+Unfortunately for us, the gradients often get smaller and smaller as the
+algorithm progresses down to the first hidden layers. As a result, the
+GD update leaves the lower layer connection weights
+virtually unchanged, and training never converges to a good
+solution. This is known in the literature as 
+the vanishing gradients problem.
+
+
    
+In other cases, the opposite can happen, namely the the gradients can grow bigger and
+bigger. The result is that many of the layers get large updates of the 
+weights the
+algorithm diverges. This is the exploding gradients problem, which is
+mostly encountered in recurrent neural networks. More generally, deep
+neural networks suffer from unstable gradients, different layers may
+learn at widely different speeds
 
 
    
 
    
@@ -291,6 +303,8 @@ logistic function in deep networks).
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs060.html b/doc/pub/NeuralNet/html/._NeuralNet-bs060.html
index 5b69507e2..4e8a17850 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs060.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs060.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,33 +255,34 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    The derivative of the Logistic funtion 
    
+
    Is the Logistic activation function (Sigmoid)  our choice? 
    
 
 
    
-Looking at the logistic activation function, when inputs become large
-(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when
-backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what
-little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so
-there is really nothing left for the lower layers.
+Although this unfortunate behavior has been empirically observed for
+quite a while (it was one of the reasons why deep neural networks were
+mostly abandoned for a long time), it is only around 2010 that
+significant progress was made in understanding it.
 
 
    
-In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the
-signal to flow properly in both directions: in the forward direction when making predictions, and in the
-reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it
-to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the
-outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have
-equal variance before and after flowing through a layer in the reverse direction (please check out the
-paper if you are interested in the mathematical details).
+A paper titled Understanding the Difficulty of Training Deep
+Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio found that
+the problems with the popular logistic
+sigmoid activation function and the weight initialization technique
+that was most popular at the time, namely random initialization using
+a normal distribution with a mean of 0 and a standard deviation of
+1.
 
 
    
-One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients
-problems were in part due to a poor choice of activation function. Until then most people had assumed
-that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they
-must be an excellent choice. But it turns out that other activation functions behave much better in deep
-neural networks, in particular the ReLU activation function, mostly because it does not saturate for
-positive values (and also because it is quite fast to compute).
+They showed that with this activation function and this
+initialization scheme, the variance of the outputs of each layer is
+much greater than the variance of its inputs. Going forward in the
+network, the variance keeps increasing after each layer until the
+activation function saturates at the top layers. This is actually made
+worse by the fact that the logistic function has a mean of 0.5, not 0
+(the hyperbolic tangent function has a mean of 0 and behaves slightly
+better than the logistic function in deep networks).
 
 
    
 
    
@@ -298,6 +303,8 @@ positive values (and also because it is quite fast to compute).
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs061.html b/doc/pub/NeuralNet/html/._NeuralNet-bs061.html
index 8bae632f7..4419076ee 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs061.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs061.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,31 +257,38 @@ MathJax.Hub.Config({
 
 
 
-
    The RELU function family 
    
+
    The derivative of the Logistic funtion 
    
 
 
    
-The ReLU activation function suffers from a problem known as the dying
-ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.
+Looking at the logistic activation function, when inputs become large
+(negative or positive), the function saturates at 0 or 1, with a
+derivative extremely close to 0. Thus when backpropagation kicks in,
+it has virtually no gradient to propagate back through the network,
+and what little gradient exists keeps getting diluted as
+backpropagation progresses down through the top layers, so there is
+really nothing left for the lower layers.
 
 
    
-In some cases, you may find that half of your network’s neurons are dead, especially if you used a large
-learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s
-inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life
-since the gradient of the ReLU function is 0 when its input is negative.
+In their paper, Glorot and Bengio propose a way to significantly
+alleviate this problem. We need the signal to flow properly in both
+directions: in the forward direction when making predictions, and in
+the reverse direction when backpropagating gradients. We don’t want
+the signal to die out, nor do we want it to explode and saturate. For
+the signal to flow properly, the authors argue that we need the
+variance of the outputs of each layer to be equal to the variance of
+its inputs, and we also need the gradients to have equal variance
+before and after flowing through a layer in the reverse direction.
 
 
    
-To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function
-$$
-ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
-$$
-
-
    
-So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,
-in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than \( \tanh \) which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you
-may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default \( \alpha \) of
-\( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have spare time and computing power, you can use
-cross-validation or bootstrap to evaluate other activation functions.
-huge training set.
+One of the insights in the 2010 paper by Glorot and Bengio was that
+the vanishing/exploding gradients problems were in part due to a poor
+choice of activation function. Until then most people had assumed that
+if Nature had chosen to use roughly sigmoid activation functions in
+biological neurons, they must be an excellent choice. But it turns out
+that other activation functions behave much better in deep neural
+networks, in particular the ReLU activation function, mostly because
+it does not saturate for positive values (and also because it is quite
+fast to compute).
 
 
    
 
    
@@ -297,6 +308,8 @@ huge training set.
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs062.html b/doc/pub/NeuralNet/html/._NeuralNet-bs062.html
index 273c0a6d2..40b629114 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs062.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs062.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -251,44 +255,31 @@ MathJax.Hub.Config({
 
     
     
     
     
 
 
-
+
 
-
    A top-down perspective on Neural networks 
    
+
    The RELU function family 
    
 
 
    
-The first thing we would like to do is divide the data into two or three
-parts. A training set, a validation or dev (development) set, and a
-test set. The test set is the data on which we want to make
-predictions. The dev set is a subset of the training data we use to
-check how well we are doing out-of-sample, after training the model on
-the training dataset. We use the validation error as a proxy for the
-test error in order to make tweaks to our model. It is crucial that we
-do not use any of the test data to train the algorithm. This is a
-cardinal sin in ML. Then:
+The ReLU activation function suffers from a problem known as the dying
+ReLUs: during training, some neurons effectively die, meaning they
+stop outputting anything other than 0.
 
-
      
-
    •  Estimate optimal error rate
      • 
-
    •  Minimize underfitting (bias) on training data set.
      • 
-
    •  Make sure you are not overfitting.
      • 
-
    
+
    
+In some cases, you may find that half of your network’s neurons are
+dead, especially if you used a large learning rate. During training,
+if a neuron’s weights get updated such that the weighted sum of the
+neuron’s inputs is negative, it will start outputting 0. When this
+happen, the neuron is unlikely to come back to life since the gradient
+of the ReLU function is 0 when its input is negative.
 
-If the validation and test sets are drawn from the same distributions,
-then good performance on the validation set should lead to similarly
-good performance on the test set. 
-However, sometimes
-the training data and test data differ in subtle ways because, for
-example, they are collected using slightly different methods, or
-because it is cheaper to collect data in one way versus another. In
-this case, there can be a mismatch between the training and test
-data. This can lead to the neural network overfitting these small
-differences between the test and training sets, and a poor performance
-on the test set despite having a good performance on the validation
-set. To rectify this, Andrew Ng suggests making two validation or dev
-sets, one constructed from the training data and one constructed from
-the test data. The difference between the performance of the algorithm
-on these two validation sets quantifies the train-test mismatch. This
-can serve as another important diagnostic when using DNNs for
-supervised learning.
+
    
+To solve this problem, nowadays practitioners use a  variant of the ReLU
+function, such as the leaky ReLU discussed above or the so-called
+exponential linear unit (ELU) function
+
+$$
+ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
+$$
 
 
    
 
    
@@ -307,6 +298,8 @@ supervised learning.
   
  • 62
    • 
   
  • 63
    • 
   
  • 64
    • 
+  
  • 65
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs063.html b/doc/pub/NeuralNet/html/._NeuralNet-bs063.html
index 28faaa4b6..8441353dd 100644
--- a/doc/pub/NeuralNet/html/._NeuralNet-bs063.html
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs063.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -253,29 +257,24 @@ MathJax.Hub.Config({
 
 
 
-
    Limitations of supervised learning with deep networks 
    
+
    Which activation function should we use? 
    
 
 
    
-Like all statistical methods, supervised learning using neural
-networks has important limitations. This is especially important when
-one seeks to apply these methods, especially to physics problems. Like
-all tools, DNNs are not a universal solution. Often, the same or
-better performance on a task can be achieved by using a few
-hand-engineered features (or even a collection of random
-features).
+In general it seems that the ELU activation function is better than
+the leaky ReLU function (and its variants), which is better than
+ReLU. ReLU performs better than \( \tanh \) which in turn performs better
+than the logistic function.
 
 
    
-Here we list some of the important limitations of supervised neural network based models. 
-
-
      
-
    •  Need labeled data. All supervised learning methods, DNNs for supervised learning require labeled data. Often, labeled data is harder to acquire than unlabeled data (e.g. one must pay for human experts to label images).
      • 
-
    •  Supervised neural networks are extremely data intensive. DNNs are data hungry. They perform best when data is plentiful. This is doubly so for supervised methods where the data must also be labeled. The utility of DNNs is extremely limited if data is hard to acquire or the datasets are small (hundreds to a few thousand samples). In this case, the performance of other methods that utilize hand-engineered features can exceed that of DNNs.
      • 
-
    •  Homogeneous data. Almost all DNNs deal with homogeneous data of one type. It is very hard to design architectures that mix and match data types (i.e. some continuous variables, some discrete variables, some time series). In applications beyond images, video, and language, this is often what is required. In contrast, ensemble models like random forests or gradient-boosted trees have no difficulty handling mixed data types.
      • 
-
    •  Many problems are not about prediction. In natural science we are often interested in learning something about the underlying distribution that generates the data. In this case, it is often difficult to cast these ideas in a supervised learning setting. While the problems are related, it is possible to make good predictions with a wrong model. The model might or might not be useful for understanding the underlying science.
      • 
-
    
-
-Some of these remarks are particular to DNNs, others are shared by all supervised learning methods. This motivates the use of unsupervised methods which in part circumnavigate these problems.
+If runtime
+performance is an issue, then you may opt for the  leaky ReLU function  over the 
+ELU function If you don’t
+want to tweak yet another hyperparameter, you may just use the default
+\( \alpha \) of \( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have
+spare time and computing power, you can use cross-validation or
+bootstrap to evaluate other activation functions.
 
+
    
 
    
 
 
      
@@ -291,6 +290,9 @@ Some of these remarks are particular to DNNs, others are shared by all supervise
   
    • 62
      • 
   
    • 63
      • 
   
    • 64
      • 
+  
    • 65
      • 
+  
    • 66
      • 
+  
    • »
      • 
 
    
 
 
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs064.html b/doc/pub/NeuralNet/html/._NeuralNet-bs064.html
new file mode 100644
index 000000000..eaae2cf1e
--- /dev/null
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs064.html
@@ -0,0 +1,340 @@
+
+
+
+
+
+
+
+Data Analysis and Machine Learning: Neural networks, from the simple perceptron to deep learning
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    
+
+
    
+  
    
+    
+    Data Analysis and Machine Learning: Neural networks, from the simple perceptron to deep learning
+  
    
+
+  
    
+    
+  
    
+
    
+
     
+
+
    
+
+
     
     
     
     
+
+
+
+
+
    A top-down perspective on Neural networks 
    
+
+
    
+The first thing we would like to do is divide the data into two or three
+parts. A training set, a validation or dev (development) set, and a
+test set. The test set is the data on which we want to make
+predictions. The dev set is a subset of the training data we use to
+check how well we are doing out-of-sample, after training the model on
+the training dataset. We use the validation error as a proxy for the
+test error in order to make tweaks to our model. It is crucial that we
+do not use any of the test data to train the algorithm. This is a
+cardinal sin in ML. Then:
+
+
      
+
    •  Estimate optimal error rate
      • 
+
    •  Minimize underfitting (bias) on training data set.
      • 
+
    •  Make sure you are not overfitting.
      • 
+
    
+
+If the validation and test sets are drawn from the same distributions,
+then a good performance on the validation set should lead to similarly
+good performance on the test set.
+
+
    
+However, sometimes
+the training data and test data differ in subtle ways because, for
+example, they are collected using slightly different methods, or
+because it is cheaper to collect data in one way versus another. In
+this case, there can be a mismatch between the training and test
+data. This can lead to the neural network overfitting these small
+differences between the test and training sets, and a poor performance
+on the test set despite having a good performance on the validation
+set. To rectify this, Andrew Ng suggests making two validation or dev
+sets, one constructed from the training data and one constructed from
+the test data. The difference between the performance of the algorithm
+on these two validation sets quantifies the train-test mismatch. This
+can serve as another important diagnostic when using DNNs for
+supervised learning.
+
+
    
+
    
+
+
    
+
+
+
      
+
+
+
+
+
+
+
+
    
+
+
    
+
+
+
+
+    
+
diff --git a/doc/pub/NeuralNet/html/._NeuralNet-bs065.html b/doc/pub/NeuralNet/html/._NeuralNet-bs065.html
new file mode 100644
index 000000000..fc43897ec
--- /dev/null
+++ b/doc/pub/NeuralNet/html/._NeuralNet-bs065.html
@@ -0,0 +1,321 @@
+
+
+
+
+
+
+
+Data Analysis and Machine Learning: Neural networks, from the simple perceptron to deep learning
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    
+
+
    
+  
    
+    
+    Data Analysis and Machine Learning: Neural networks, from the simple perceptron to deep learning
+  
    
+
+  
    
+    
+  
    
+
    
+
     
+
+
    
+
+
     
     
     
     
+
+
+
+
+
    Limitations of supervised learning with deep networks 
    
+
+
    
+Like all statistical methods, supervised learning using neural
+networks has important limitations. This is especially important when
+one seeks to apply these methods, especially to physics problems. Like
+all tools, DNNs are not a universal solution. Often, the same or
+better performance on a task can be achieved by using a few
+hand-engineered features (or even a collection of random
+features).
+
+
    
+Here we list some of the important limitations of supervised neural network based models. 
+
+
      
+
    •  Need labeled data. All supervised learning methods, DNNs for supervised learning require labeled data. Often, labeled data is harder to acquire than unlabeled data (e.g. one must pay for human experts to label images).
      • 
+
    •  Supervised neural networks are extremely data intensive. DNNs are data hungry. They perform best when data is plentiful. This is doubly so for supervised methods where the data must also be labeled. The utility of DNNs is extremely limited if data is hard to acquire or the datasets are small (hundreds to a few thousand samples). In this case, the performance of other methods that utilize hand-engineered features can exceed that of DNNs.
      • 
+
    •  Homogeneous data. Almost all DNNs deal with homogeneous data of one type. It is very hard to design architectures that mix and match data types (i.e. some continuous variables, some discrete variables, some time series). In applications beyond images, video, and language, this is often what is required. In contrast, ensemble models like random forests or gradient-boosted trees have no difficulty handling mixed data types.
      • 
+
    •  Many problems are not about prediction. In natural science we are often interested in learning something about the underlying distribution that generates the data. In this case, it is often difficult to cast these ideas in a supervised learning setting. While the problems are related, it is possible to make good predictions with a wrong model. The model might or might not be useful for understanding the underlying science.
      • 
+
    
+
+Some of these remarks are particular to DNNs, others are shared by all supervised learning methods. This motivates the use of unsupervised methods which in part circumnavigate these problems.
+
+
    
+
+
    
+
+
+
      
+
+
+
+
+
+
+
+
    
+
+
    
+
+
+
+
+    
+
diff --git a/doc/pub/NeuralNet/html/NeuralNet-bs.html b/doc/pub/NeuralNet/html/NeuralNet-bs.html
index e3971cad4..aeabd57ad 100644
--- a/doc/pub/NeuralNet/html/NeuralNet-bs.html
+++ b/doc/pub/NeuralNet/html/NeuralNet-bs.html
@@ -91,53 +91,55 @@ Automatically generated HTML file from DocOnce source
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -207,37 +209,39 @@ MathJax.Hub.Config({
       
  • Setting up the Back propagation algorithm
    • 
       
  • Setting up a Multi-layer perceptron model for classification
    • 
       
  • Defining the cost function
    • 
-      
  • Developing a code for doing neural networks with back propagation
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Train and test datasets
    • 
-      
  • Define model and architecture
    • 
-      
  • Layers
    • 
-      
  • Weights and biases
    • 
-      
  • Feed-forward pass
    • 
-      
  • Matrix multiplications
    • 
-      
  • Choose cost function and optimizer
    • 
-      
  • Optimizing the cost function
    • 
-      
  • Regularization
    • 
-      
  • Matrix  multiplication
    • 
-      
  • Improving performance
    • 
-      
  • Full object-oriented implementation
    • 
-      
  • Evaluate model performance on test data
    • 
-      
  • Adjust hyperparameters
    • 
-      
  • Visualization
    • 
-      
  • scikit-learn implementation
    • 
-      
  • Visualization
    • 
-      
  • Building neural networks in Tensorflow and Keras
    • 
-      
  • Tensorflow
    • 
-      
  • Collect and pre-process data
    • 
-      
  • Using TensorFlow backend
    • 
-      
  • Optimizing and using gradient descent
    • 
-      
  • Using Keras
    • 
-      
  • Which activation function should I use?
    • 
-      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
-      
  • The derivative of the Logistic funtion
    • 
-      
  • The RELU function family
    • 
-      
  • A top-down perspective on Neural networks
    • 
-      
  • Limitations of supervised learning with deep networks
    • 
+      
  • Example: binary classification problem
    • 
+      
  • Developing a code for doing neural networks with back propagation
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Train and test datasets
    • 
+      
  • Define model and architecture
    • 
+      
  • Layers
    • 
+      
  • Weights and biases
    • 
+      
  • Feed-forward pass
    • 
+      
  • Matrix multiplications
    • 
+      
  • Choose cost function and optimizer
    • 
+      
  • Optimizing the cost function
    • 
+      
  • Regularization
    • 
+      
  • Matrix  multiplication
    • 
+      
  • Improving performance
    • 
+      
  • Full object-oriented implementation
    • 
+      
  • Evaluate model performance on test data
    • 
+      
  • Adjust hyperparameters
    • 
+      
  • Visualization
    • 
+      
  • scikit-learn implementation
    • 
+      
  • Visualization
    • 
+      
  • Building neural networks in Tensorflow and Keras
    • 
+      
  • Tensorflow
    • 
+      
  • Collect and pre-process data
    • 
+      
  • Using TensorFlow backend
    • 
+      
  • Optimizing and using gradient descent
    • 
+      
  • Using Keras
    • 
+      
  • Which activation function should I use?
    • 
+      
  • Is the Logistic activation function (Sigmoid)  our choice?
    • 
+      
  • The derivative of the Logistic funtion
    • 
+      
  • The RELU function family
    • 
+      
  • Which activation function should we use?
    • 
+      
  • A top-down perspective on Neural networks
    • 
+      
  • Limitations of supervised learning with deep networks
    • 
 
         
       
@@ -272,7 +276,7 @@ MathJax.Hub.Config({
 
    [2] Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University
    
 
    
 
    
-
    Oct 12, 2018
     
+
    Oct 16, 2018
     
 
    
 
    
 
@@ -296,7 +300,7 @@ MathJax.Hub.Config({
   
  • 9
    • 
   
  • 10
    • 
   
  • ...
    • 
-  
  • 64
    • 
+  
  • 66
    • 
   
  • »
    • 
 
 
diff --git a/doc/pub/NeuralNet/html/NeuralNet-reveal.html b/doc/pub/NeuralNet/html/NeuralNet-reveal.html
index 1022e6c03..59ebbaf6a 100644
--- a/doc/pub/NeuralNet/html/NeuralNet-reveal.html
+++ b/doc/pub/NeuralNet/html/NeuralNet-reveal.html
@@ -148,7 +148,7 @@ MathJax.Hub.Config({
 
    [2] Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University
    
 
    
 
     
    
-
    Oct 12, 2018
     
+
    Oct 16, 2018
     
 
    
 
    
 
@@ -1123,7 +1123,7 @@ z_j^{l+1} = \sum_{i=1}^{M_{l}}w_{ij}^{l+1}a_j^{l}+b_j^{l+1},
 $$
 
     
    
 
-we obtain
+with \( M_l \) being the number of nodes in layer \( l \), we obtain
 
     
    
 $$
 \delta_j^l =\sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l),
@@ -1181,7 +1181,7 @@ $$
 Then we compute the back propagate error for each \( l=L-1,L-2,\dots,2 \) as
 
     
    
 $$
-\delta_j^l =\sum_k \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
+\delta_j^l = \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
 $$
 
     
    
 
    
@@ -1190,7 +1190,7 @@ $$
 
    
 
 
    
-Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,dots,2 \) and update the weights and biases according to the rules
+Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,\dots,2 \) and update the weights and biases according to the rules
 
     
    
 $$
 w_{jk}^l\leftarrow  = w_{jk}^l- \eta \delta_j^la_k^{l-1},
@@ -1199,7 +1199,7 @@ $$
 
 
     
    
 $$
-b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^L},
+b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^l}=b_j^l-\eta \delta_j^l,
 $$
 
     
    
 
    
@@ -1309,7 +1309,7 @@ Again we take the negative log-likelihood to define our cost function:
 
 
     
    
 $$
-\mathcal{C}(\hat{\theta}) = - \ln P(\mathcal{D} \mid \hat{\theta}).
+\mathcal{C}(\hat{\theta}) = - \log{P(\mathcal{D} \mid \hat{\theta})}.
 $$
 
     
    
 
@@ -1317,12 +1317,73 @@ See the logistic regression lectures for a full definition of the cost function.
 
 
    
 The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before!
-We leave it as an exercise in project 2 to derive these equations.
 
 
 
 
    
-
    Developing a code for doing neural networks with back propagation 
    
+
    Example: binary classification problem 
    
+
+
    
+As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters \( \beta \) as
+
     
    
+$$
+\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right),
+$$
+
     
    
+
+where we had defined the logistic (sigmoid) function
+
     
    
+$$
+p(y_i =1\vert x_i,\hat{\beta})=\frac{\exp{(\beta_0+\beta_1 x_i)}}{1+\exp{(\beta_0+\beta_1 x_i)}},
+$$
+
     
    
+
+and
+
     
    
+$$
+p(y_i =0\vert x_i,\hat{\beta})=1-p(y_i =1\vert x_i,\hat{\beta}).
+$$
+
     
    
+
+The parameters \( \hat{\beta} \) were defined using a minimization method like gradient descent or Newton-Raphson's method.
+
+
    
+Now we replace \( x_i \) with the activation \( z_i^l \) for a given layer \( l \) and the outputs as \( y_i=a_i^l=f(z_i^l) \), with \( z_i^l \) now being a function of the weights \( w_{ij}^l \) and biases \( b_i^l \). 
+We have then
+
     
    
+$$
+a_i^l = y_i = \frac{\exp{(z_i^l)}}{1+\exp{(z_i^l)}},
+$$
+
     
    
+
+with 
+
     
    
+$$
+z_i^l = \sum_{j}w_{ij}^l a_j^{l-1}+b_i^l,
+$$
+
     
    
+
+where the superscript \( l-1 \) indicates that these are the outputs from layer \( l-1 \).
+Our cost function at the final layer \( l=L \) is now
+
     
    
+$$
+\mathcal{C}(\hat{W}) = - \sum_{i=1}^n \left(t_i\log{a_i^L}+(i-t_i)\log{(1-a_i^L)}\right),
+$$
+
     
    
+
+where we have defined the targets \( t_i \). The derivatives of the cost function with respect to the output \( a_i^L \) are then easily calculated and we get
+
     
    
+$$
+\frac{\partial \mathcal{C}(\hat{W})}{\partial a_i^L} = \frac{a_i^L-t_i}{a_i^L(1-a_i^L)}. 
+$$
+
     
    
+
+In case we use another activation function than the logistic one, we need to evaluate other derivatives.
+
    
+
+
+
    
+
    Developing a code for doing neural networks with back propagation 
    
 
 
    
 One can identify a set of key steps when using neural networks to solve supervised learning problems:  
@@ -1344,7 +1405,7 @@ One can identify a set of key steps when using neural networks to solve supervis
 
 
 
    
-
    Collect and pre-process data 
    
+
    Collect and pre-process data 
    
 
 
    
 Here we will be using the MNIST dataset, which is readily available through the scikit-learn
@@ -1444,7 +1505,7 @@ plt.show()
 
 
 
    
-
    Train and test datasets 
    
+
    Train and test datasets 
    
 
 
    
 Performing analysis before partitioning the dataset is a major error, that can lead to incorrect conclusions.
@@ -1494,7 +1555,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
 
 
    
-
    Define model and architecture 
    
+
    Define model and architecture 
    
 
 
    
 Our simple feed-forward neural network will consist of an input layer, a single hidden layer and an output layer. The activation \( y \) of each neuron is a weighted sum of inputs, passed through an activation function. In case of the simple perceptron model we have
@@ -1543,12 +1604,12 @@ $$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
 
     
    
 
 
    
-which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.
+which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.
 
    
 
 
 
    
-
    Layers 
    
+
    Layers 
    
 
 
      
 
    •  Input
      •  
@@ -1600,7 +1661,7 @@ weights to the output layer.
 
 
 
      
-
      Weights and biases 
      
+
      Weights and biases 
      
 
 
      
 Typically weights are initialized with small values distributed around zero, drawn from a uniform
@@ -1639,7 +1700,7 @@ output_bias = np.zeros(n_categories) + 0.01
 
 
 
      
-
      Feed-forward pass 
      
+
      Feed-forward pass 
      
 
 
      
 Denote \( F \) the number of features, \( H \) the number of hidden neurons and \( C \) the number of categories.  
@@ -1674,7 +1735,7 @@ $$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
 
 
 
      
-
      Matrix multiplications 
      
+
      Matrix multiplications 
      
 
 
      
 Since our data has the dimensions \( X = (n_{inputs}, n_{features}) \) and our weights to the hidden
@@ -1760,7 +1821,7 @@ predictions = predict(X_train)
 
 
 
      
-
      Choose cost function and optimizer 
      
+
      Choose cost function and optimizer 
      
 
 
      
 To measure how well our neural network is doing we need to introduce a cost function.  
@@ -1795,7 +1856,7 @@ you got the correct label. The probability of category \( c \) is given by the s
 
 
 
      
-
      Optimizing the cost function 
      
+
      Optimizing the cost function 
      
 
 
      
 The network is trained by finding the weights and biases that minimize the cost function. One of the most widely used classes of methods is gradient descent and its generalizations. The idea behind gradient descent
@@ -1834,11 +1895,14 @@ This has two important benefits:
 
 
    •  It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
      •   
 
+
      
+
+The various optmization  methods, with codes and algorithms,  are discussed in our lectures on Gradient descent approaches.
 
      
 
 
 
      
-
      Regularization 
      
+
      Regularization 
      
 
 
      
 It is common to add an extra term to the cost function, proportional
@@ -1873,14 +1937,14 @@ calculate the gradient efficently.
 
 
 
      
-
      Matrix  multiplication 
      
+
      Matrix  multiplication 
      
 
 
      
 To more efficently train our network these equations are implemented using matrix operations.  
-The error in the output layer is calculated simply as
+The error in the output layer is calculated simply as, with \( \hat{t} \) being our targets,
 
 
       
      
-$$ \delta_L = \hat{y} - y = (n_{inputs}, n_{categories}) .$$
+$$ \delta_L = \hat{t} - \hat{y} = (n_{inputs}, n_{categories}) .$$
 
       
      
 
 
      
@@ -1928,9 +1992,6 @@ $$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
 
 
 
      # to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our model
 from sklearn.metrics import accuracy_score
 
 # one-hot in numpy
@@ -2003,7 +2064,7 @@ lmbd = 0.01
 
 
 
      
-
      Improving performance 
      
+
      Improving performance 
      
 
 
      
 As we can see the network does not seem to be learning at all. It seems to be just guessing the label for each image.  
@@ -2023,7 +2084,7 @@ Andrew Ng goes through some of these considerations in this Full object-oriented implementation 
+
      Full object-oriented implementation 
      
 
 
      
 It is very natural to think of the network as an object, with specific instances of the network
@@ -2137,7 +2198,7 @@ being realizations of this object with different hyperparameters. An implementat
 
 
 
      
-
      Evaluate model performance on test data 
      
+
      Evaluate model performance on test data 
      
 
 
      
 To measure the performance of our network we evaluate how well it does it data it has never seen before, i.e. the test data.  
@@ -2175,7 +2236,7 @@ test_predict = dnn.predict(X_test)
 
 
 
      
-
      Adjust hyperparameters 
      
+
      Adjust hyperparameters 
      
 
 
      
 We now perform a grid search to find the optimal hyperparameters for the network.  
@@ -2209,7 +2270,7 @@ DNN_numpy = np.zeros((len(eta_vals), 
-
      Visualization 
      
+
      Visualization 
      
 
 
      
 
@@ -2252,7 +2313,7 @@ plt.show()
 
 
 
      
-
      scikit-learn implementation 
      
+
      scikit-learn implementation 
      
 
 
      
 scikit-learn focuses more
@@ -2292,7 +2353,7 @@ DNN_scikit = np.zeros((len(eta_vals), 
-
      Visualization 
      
+
      Visualization 
      
 
      
 
 
@@ -2335,7 +2396,7 @@ plt.show()
 
 
 
      
-
      Building neural networks in Tensorflow and Keras 
      
+
      Building neural networks in Tensorflow and Keras 
      
 
 
      
 Now we want  to build on the experience gained from our neural network implementation in NumPy and scikit-learn
@@ -2350,7 +2411,7 @@ NumPy arrays.
 
 
 
      
-
      Tensorflow 
      
+
      Tensorflow 
      
 
 
      
 Tensorflow is an open source library machine learning library
@@ -2396,7 +2457,7 @@ and/or if you use anaconda, just write (or install from the graphical use
 
 
 
      
-
      Collect and pre-process data 
      
+
      Collect and pre-process data 
      
 
 
      
 
@@ -2463,7 +2524,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
 
 
      
-
      Using TensorFlow backend 
      
+
      Using TensorFlow backend 
      
 
 
        
 
      1.  Define model and architecture
        2. 
@@ -2608,7 +2669,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
 
 
        
-
        Optimizing and using gradient descent 
        
+
        Optimizing and using gradient descent 
        
 
 
        
 
@@ -2687,7 +2748,7 @@ writer.add_graph(tf.get_default_graph())
 
 
 
        
-
        Using Keras 
        
+
        Using Keras 
        
 
 
        
 Keras is a high level neural network
@@ -2787,108 +2848,151 @@ plt.show()
 
 
 
        
-
        Which activation function should I use? 
        
+
        Which activation function should I use? 
        
 
 
        
-Backpropagation algorithm works by going from the output layer to the
-input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the
-cost function with regards to each parameter in the network, it uses these gradients to update each
-parameter with a Gradient Descent step.
+The Back propagation algorithm we derived above works by going from
+the output layer to the input layer, propagating the error gradient on
+the way. Once the algorithm has computed the gradient of the cost
+function with regards to each parameter in the network, it uses these
+gradients to update each parameter with a Gradient Descent (GD) step.
 
 
        
-Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower
-layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually
-unchanged, and training never converges to a good solution. This is called the vanishing gradients
-problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many
-layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients
-problem, which is mostly encountered in recurrent neural networks. More generally,
-deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds
+Unfortunately for us, the gradients often get smaller and smaller as the
+algorithm progresses down to the first hidden layers. As a result, the
+GD update leaves the lower layer connection weights
+virtually unchanged, and training never converges to a good
+solution. This is known in the literature as 
+the vanishing gradients problem.
+
+
        
+In other cases, the opposite can happen, namely the the gradients can grow bigger and
+bigger. The result is that many of the layers get large updates of the 
+weights the
+algorithm diverges. This is the exploding gradients problem, which is
+mostly encountered in recurrent neural networks. More generally, deep
+neural networks suffer from unstable gradients, different layers may
+learn at widely different speeds
 
        
 
 
 
        
-
        Is the Logistic activation function (Sigmoid)  our choice? 
        
+
        Is the Logistic activation function (Sigmoid)  our choice? 
        
 
 
        
-Although this unfortunate behavior has been empirically observed for quite a while (it was one of the
-reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that
+Although this unfortunate behavior has been empirically observed for
+quite a while (it was one of the reasons why deep neural networks were
+mostly abandoned for a long time), it is only around 2010 that
 significant progress was made in understanding it.
 
 
        
-A paper titled Understanding the Difficulty of Training Deep Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio1 found a few suspects,
-including the combination of the popular logistic sigmoid activation function and the weight initialization
-technique that was most popular at the time, namely random initialization using a normal distribution with
-a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this
-initialization scheme, the variance of the outputs of each layer is much greater than the variance of its
-inputs. Going forward in the network, the variance keeps increasing after each layer until the activation
-function saturates at the top layers. This is actually made worse by the fact that the logistic function has a
-mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the
-logistic function in deep networks).
+A paper titled Understanding the Difficulty of Training Deep
+Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio found that
+the problems with the popular logistic
+sigmoid activation function and the weight initialization technique
+that was most popular at the time, namely random initialization using
+a normal distribution with a mean of 0 and a standard deviation of
+1.
+
+
        
+They showed that with this activation function and this
+initialization scheme, the variance of the outputs of each layer is
+much greater than the variance of its inputs. Going forward in the
+network, the variance keeps increasing after each layer until the
+activation function saturates at the top layers. This is actually made
+worse by the fact that the logistic function has a mean of 0.5, not 0
+(the hyperbolic tangent function has a mean of 0 and behaves slightly
+better than the logistic function in deep networks).
 
        
 
 
 
        
-
        The derivative of the Logistic funtion 
        
+
        The derivative of the Logistic funtion 
        
 
 
        
 Looking at the logistic activation function, when inputs become large
-(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when
-backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what
-little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so
-there is really nothing left for the lower layers.
+(negative or positive), the function saturates at 0 or 1, with a
+derivative extremely close to 0. Thus when backpropagation kicks in,
+it has virtually no gradient to propagate back through the network,
+and what little gradient exists keeps getting diluted as
+backpropagation progresses down through the top layers, so there is
+really nothing left for the lower layers.
 
 
        
-In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the
-signal to flow properly in both directions: in the forward direction when making predictions, and in the
-reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it
-to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the
-outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have
-equal variance before and after flowing through a layer in the reverse direction (please check out the
-paper if you are interested in the mathematical details).
+In their paper, Glorot and Bengio propose a way to significantly
+alleviate this problem. We need the signal to flow properly in both
+directions: in the forward direction when making predictions, and in
+the reverse direction when backpropagating gradients. We don’t want
+the signal to die out, nor do we want it to explode and saturate. For
+the signal to flow properly, the authors argue that we need the
+variance of the outputs of each layer to be equal to the variance of
+its inputs, and we also need the gradients to have equal variance
+before and after flowing through a layer in the reverse direction.
 
 
        
-One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients
-problems were in part due to a poor choice of activation function. Until then most people had assumed
-that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they
-must be an excellent choice. But it turns out that other activation functions behave much better in deep
-neural networks, in particular the ReLU activation function, mostly because it does not saturate for
-positive values (and also because it is quite fast to compute).
+One of the insights in the 2010 paper by Glorot and Bengio was that
+the vanishing/exploding gradients problems were in part due to a poor
+choice of activation function. Until then most people had assumed that
+if Nature had chosen to use roughly sigmoid activation functions in
+biological neurons, they must be an excellent choice. But it turns out
+that other activation functions behave much better in deep neural
+networks, in particular the ReLU activation function, mostly because
+it does not saturate for positive values (and also because it is quite
+fast to compute).
 
        
 
 
 
        
-
        The RELU function family 
        
+
        The RELU function family 
        
 
 
        
 The ReLU activation function suffers from a problem known as the dying
-ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.
+ReLUs: during training, some neurons effectively die, meaning they
+stop outputting anything other than 0.
 
 
        
-In some cases, you may find that half of your network’s neurons are dead, especially if you used a large
-learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s
-inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life
-since the gradient of the ReLU function is 0 when its input is negative.
+In some cases, you may find that half of your network’s neurons are
+dead, especially if you used a large learning rate. During training,
+if a neuron’s weights get updated such that the weighted sum of the
+neuron’s inputs is negative, it will start outputting 0. When this
+happen, the neuron is unlikely to come back to life since the gradient
+of the ReLU function is 0 when its input is negative.
 
 
        
-To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function
+To solve this problem, nowadays practitioners use a  variant of the ReLU
+function, such as the leaky ReLU discussed above or the so-called
+exponential linear unit (ELU) function
+
 
         
        
 $$
 ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
 $$
 
         
        
-
-
        
-So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,
-in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than \( \tanh \) which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you
-may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default \( \alpha \) of
-\( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have spare time and computing power, you can use
-cross-validation or bootstrap to evaluate other activation functions.
-huge training set.
 
        
 
 
 
        
-
        A top-down perspective on Neural networks 
        
+
        Which activation function should we use? 
        
+
+
        
+In general it seems that the ELU activation function is better than
+the leaky ReLU function (and its variants), which is better than
+ReLU. ReLU performs better than \( \tanh \) which in turn performs better
+than the logistic function.
+
+
        
+If runtime
+performance is an issue, then you may opt for the  leaky ReLU function  over the 
+ELU function If you don’t
+want to tweak yet another hyperparameter, you may just use the default
+\( \alpha \) of \( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have
+spare time and computing power, you can use cross-validation or
+bootstrap to evaluate other activation functions.
+
        
+
+
+
        
+
        A top-down perspective on Neural networks 
        
 
 
        
 The first thing we would like to do is divide the data into two or three
@@ -2909,8 +3013,10 @@ cardinal sin in ML. Then:
 
        
 
 If the validation and test sets are drawn from the same distributions,
-then good performance on the validation set should lead to similarly
-good performance on the test set. 
+then a good performance on the validation set should lead to similarly
+good performance on the test set.
+
+
        
 However, sometimes
 the training data and test data differ in subtle ways because, for
 example, they are collected using slightly different methods, or
@@ -2929,7 +3035,7 @@ supervised learning.
 
 
 
        
-
        Limitations of supervised learning with deep networks 
        
+
        Limitations of supervised learning with deep networks 
        
 
 
        
 Like all statistical methods, supervised learning using neural
diff --git a/doc/pub/NeuralNet/html/NeuralNet-solarized.html b/doc/pub/NeuralNet/html/NeuralNet-solarized.html
index 5aa97c64b..fa2aa10d3 100644
--- a/doc/pub/NeuralNet/html/NeuralNet-solarized.html
+++ b/doc/pub/NeuralNet/html/NeuralNet-solarized.html
@@ -111,53 +111,55 @@ div { text-align: justify; text-justify: inter-word; }
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -199,7 +201,7 @@ MathJax.Hub.Config({
 
        [2] Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University
        
 
        
 
        
-
        Oct 12, 2018
         
+
        Oct 16, 2018
         
 
        
 
        
 









        
@@ -1084,7 +1086,7 @@ $$
 z_j^{l+1} = \sum_{i=1}^{M_{l}}w_{ij}^{l+1}a_j^{l}+b_j^{l+1},
 $$
 
-we obtain
+with \( M_l \) being the number of nodes in layer \( l \), we obtain
 $$
 \delta_j^l =\sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l),
 $$
@@ -1140,7 +1142,7 @@ $$
 
        
 Then we compute the back propagate error for each \( l=L-1,L-2,\dots,2 \) as
 $$
-\delta_j^l =\sum_k \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
+\delta_j^l = \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
 $$
 
      
 
@@ -1149,14 +1151,14 @@ $$
 
      
 
 
      
-Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,dots,2 \) and update the weights and biases according to the rules
+Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,\dots,2 \) and update the weights and biases according to the rules
 $$
 w_{jk}^l\leftarrow  = w_{jk}^l- \eta \delta_j^la_k^{l-1},
 $$
 
 
 $$
-b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^L},
+b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^l}=b_j^l-\eta \delta_j^l,
 $$
 
      
 
@@ -1255,19 +1257,66 @@ $$
 Again we take the negative log-likelihood to define our cost function:  
 
 $$
-\mathcal{C}(\hat{\theta}) = - \ln P(\mathcal{D} \mid \hat{\theta}).
+\mathcal{C}(\hat{\theta}) = - \log{P(\mathcal{D} \mid \hat{\theta})}.
 $$
 
 See the logistic regression lectures for a full definition of the cost function.
 
 
      
 The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before!
-We leave it as an exercise in project 2 to derive these equations.
+
+
      
+









      
+
+
      Example: binary classification problem 
      
+
+
      
+As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters \( \beta \) as
+$$
+\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right),
+$$
+
+where we had defined the logistic (sigmoid) function
+$$
+p(y_i =1\vert x_i,\hat{\beta})=\frac{\exp{(\beta_0+\beta_1 x_i)}}{1+\exp{(\beta_0+\beta_1 x_i)}},
+$$
+
+and
+$$
+p(y_i =0\vert x_i,\hat{\beta})=1-p(y_i =1\vert x_i,\hat{\beta}).
+$$
+
+The parameters \( \hat{\beta} \) were defined using a minimization method like gradient descent or Newton-Raphson's method.
+
+
      
+Now we replace \( x_i \) with the activation \( z_i^l \) for a given layer \( l \) and the outputs as \( y_i=a_i^l=f(z_i^l) \), with \( z_i^l \) now being a function of the weights \( w_{ij}^l \) and biases \( b_i^l \). 
+We have then
+$$
+a_i^l = y_i = \frac{\exp{(z_i^l)}}{1+\exp{(z_i^l)}},
+$$
+
+with 
+$$
+z_i^l = \sum_{j}w_{ij}^l a_j^{l-1}+b_i^l,
+$$
+
+where the superscript \( l-1 \) indicates that these are the outputs from layer \( l-1 \).
+Our cost function at the final layer \( l=L \) is now
+$$
+\mathcal{C}(\hat{W}) = - \sum_{i=1}^n \left(t_i\log{a_i^L}+(i-t_i)\log{(1-a_i^L)}\right),
+$$
+
+where we have defined the targets \( t_i \). The derivatives of the cost function with respect to the output \( a_i^L \) are then easily calculated and we get
+$$
+\frac{\partial \mathcal{C}(\hat{W})}{\partial a_i^L} = \frac{a_i^L-t_i}{a_i^L(1-a_i^L)}. 
+$$
+
+In case we use another activation function than the logistic one, we need to evaluate other derivatives.
 
 
      
 
 
-
      Developing a code for doing neural networks with back propagation 
      
+
      Developing a code for doing neural networks with back propagation 
      
 
 
      
 One can identify a set of key steps when using neural networks to solve supervised learning problems:  
@@ -1283,7 +1332,7 @@ One can identify a set of key steps when using neural networks to solve supervis
 
 









      
 
-
      Collect and pre-process data 
      
+
      Collect and pre-process data 
      
 
 
      
 Here we will be using the MNIST dataset, which is readily available through the scikit-learn
@@ -1378,7 +1427,7 @@ plt.show()
 
      
 









      
 
-
      Train and test datasets 
      
+
      Train and test datasets 
      
 
 
      
 Performing analysis before partitioning the dataset is a major error, that can lead to incorrect conclusions.
@@ -1427,7 +1476,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
      
 









      
 
-
      Define model and architecture 
      
+
      Define model and architecture 
      
 
 
      
 Our simple feed-forward neural network will consist of an input layer, a single hidden layer and an output layer. The activation \( y \) of each neuron is a weighted sum of inputs, passed through an activation function. In case of the simple perceptron model we have 
@@ -1468,12 +1517,12 @@ We will be using the sigmoid function \( \sigma(x) \):
 $$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
 
 
      
-which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.
+which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.
 
 
      
 
 
-
      Layers 
      
+
      Layers 
      
 
 
        
 
      •  Input
        •  
@@ -1518,7 +1567,7 @@ weights to the output layer.
 
        
 
 
-
        Weights and biases 
        
+
        Weights and biases 
        
 
 
        
 Typically weights are initialized with small values distributed around zero, drawn from a uniform
@@ -1554,7 +1603,7 @@ output_bias = np.zeros(n_categories) + 0.01
 
        
 









        
 
-
        Feed-forward pass 
        
+
        Feed-forward pass 
        
 
 
        
 Denote \( F \) the number of features, \( H \) the number of hidden neurons and \( C \) the number of categories.  
@@ -1581,7 +1630,7 @@ $$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
 
        
 
 
-
        Matrix multiplications 
        
+
        Matrix multiplications 
        
 
 
        
 Since our data has the dimensions \( X = (n_{inputs}, n_{features}) \) and our weights to the hidden
@@ -1656,7 +1705,7 @@ predictions = predict(X_train)
 
        
 









        
 
-
        Choose cost function and optimizer 
        
+
        Choose cost function and optimizer 
        
 
 
        
 To measure how well our neural network is doing we need to introduce a cost function.  
@@ -1687,7 +1736,7 @@ you got the correct label. The probability of category \( c \) is given by the s
 
        
 









        
 
-
        Optimizing the cost function 
        
+
        Optimizing the cost function 
        
 
 
        
 The network is trained by finding the weights and biases that minimize the cost function. One of the most widely used classes of methods is gradient descent and its generalizations. The idea behind gradient descent
@@ -1722,9 +1771,12 @@ This has two important benefits:
 
      •  It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
        •   
 
 
+The various optmization  methods, with codes and algorithms,  are discussed in our lectures on Gradient descent approaches.
+
+
        
 
 
-
        Regularization 
        
+
        Regularization 
        
 
 
        
 It is common to add an extra term to the cost function, proportional
@@ -1757,13 +1809,13 @@ calculate the gradient efficently.
 
        
 









        
 
-
        Matrix  multiplication 
        
+
        Matrix  multiplication 
        
 
 
        
 To more efficently train our network these equations are implemented using matrix operations.  
-The error in the output layer is calculated simply as  
+The error in the output layer is calculated simply as, with \( \hat{t} \) being our targets,  
 
-$$ \delta_L = \hat{y} - y = (n_{inputs}, n_{categories}) .$$
+$$ \delta_L = \hat{t} - \hat{y} = (n_{inputs}, n_{categories}) .$$
 
 
        
 The gradient for the output weights is calculated as  
@@ -1800,9 +1852,6 @@ $$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
 
 
 
        # to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our model
 from sklearn.metrics import accuracy_score
 
 # one-hot in numpy
@@ -1874,7 +1923,7 @@ lmbd = 0.01
 
        
 









        
 
-
        Improving performance 
        
+
        Improving performance 
        
 
 
        
 As we can see the network does not seem to be learning at all. It seems to be just guessing the label for each image.  
@@ -1894,7 +1943,7 @@ Andrew Ng goes through some of these considerations in this Full object-oriented implementation 
+
        Full object-oriented implementation 
        
 
 
        
 It is very natural to think of the network as an object, with specific instances of the network
@@ -2007,7 +2056,7 @@ being realizations of this object with different hyperparameters. An implementat
 
        
 









        
 
-
        Evaluate model performance on test data 
        
+
        Evaluate model performance on test data 
        
 
 
        
 To measure the performance of our network we evaluate how well it does it data it has never seen before, i.e. the test data.  
@@ -2042,7 +2091,7 @@ test_predict = dnn.predict(X_test)
 
        
 









        
 
-
        Adjust hyperparameters 
        
+
        Adjust hyperparameters 
        
 
 
        
 We now perform a grid search to find the optimal hyperparameters for the network.  
@@ -2075,7 +2124,7 @@ DNN_numpy = np.zeros((len(eta_vals), 
 









        
 
-
        Visualization 
        
+
        Visualization 
        
 
 
        
 
@@ -2117,7 +2166,7 @@ plt.show()
 
        
 









        
 
-
        scikit-learn implementation 
        
+
        scikit-learn implementation 
        
 
 
        
 scikit-learn focuses more
@@ -2156,7 +2205,7 @@ DNN_scikit = np.zeros((len(eta_vals), 
 









        
 
-
        Visualization 
        
+
        Visualization 
        
 
        
 
 
@@ -2198,7 +2247,7 @@ plt.show()
 
        
 









        
 
-
        Building neural networks in Tensorflow and Keras 
        
+
        Building neural networks in Tensorflow and Keras 
        
 
 
        
 Now we want  to build on the experience gained from our neural network implementation in NumPy and scikit-learn
@@ -2213,7 +2262,7 @@ NumPy arrays.
 
        
 









        
 
-
        Tensorflow 
        
+
        Tensorflow 
        
 
 
        
 Tensorflow is an open source library machine learning library
@@ -2258,7 +2307,7 @@ and/or if you use anaconda, just write (or install from the graphical use
 
        
 









        
 
-
        Collect and pre-process data 
        
+
        Collect and pre-process data 
        
 
 
        
 
@@ -2324,7 +2373,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
        
 









        
 
-
        Using TensorFlow backend 
        
+
        Using TensorFlow backend 
        
 
 
          
 
        1.  Define model and architecture
          2. 
@@ -2469,7 +2518,7 @@ X_train, X_test, Y_train, Y_test = train_test_split(inputs, labels, train_size=t
 
          
 









          
 
-
          Optimizing and using gradient descent 
          
+
          Optimizing and using gradient descent 
          
 
 
          
 
@@ -2547,7 +2596,7 @@ writer.add_graph(tf.get_default_graph())
 
          
 









          
 
-
          Using Keras 
          
+
          Using Keras 
          
 
 
          
 Keras is a high level neural network
@@ -2646,106 +2695,149 @@ plt.show()
 
          
 
 
-
          Which activation function should I use? 
          
+
          Which activation function should I use? 
          
 
 
          
-Backpropagation algorithm works by going from the output layer to the
-input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the
-cost function with regards to each parameter in the network, it uses these gradients to update each
-parameter with a Gradient Descent step.
+The Back propagation algorithm we derived above works by going from
+the output layer to the input layer, propagating the error gradient on
+the way. Once the algorithm has computed the gradient of the cost
+function with regards to each parameter in the network, it uses these
+gradients to update each parameter with a Gradient Descent (GD) step.
 
 
          
-Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower
-layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually
-unchanged, and training never converges to a good solution. This is called the vanishing gradients
-problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many
-layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients
-problem, which is mostly encountered in recurrent neural networks. More generally,
-deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds
+Unfortunately for us, the gradients often get smaller and smaller as the
+algorithm progresses down to the first hidden layers. As a result, the
+GD update leaves the lower layer connection weights
+virtually unchanged, and training never converges to a good
+solution. This is known in the literature as 
+the vanishing gradients problem.
+
+
          
+In other cases, the opposite can happen, namely the the gradients can grow bigger and
+bigger. The result is that many of the layers get large updates of the 
+weights the
+algorithm diverges. This is the exploding gradients problem, which is
+mostly encountered in recurrent neural networks. More generally, deep
+neural networks suffer from unstable gradients, different layers may
+learn at widely different speeds
 
 
          
 
 
-
          Is the Logistic activation function (Sigmoid)  our choice? 
          
+
          Is the Logistic activation function (Sigmoid)  our choice? 
          
 
 
          
-Although this unfortunate behavior has been empirically observed for quite a while (it was one of the
-reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that
+Although this unfortunate behavior has been empirically observed for
+quite a while (it was one of the reasons why deep neural networks were
+mostly abandoned for a long time), it is only around 2010 that
 significant progress was made in understanding it.
 
 
          
-A paper titled Understanding the Difficulty of Training Deep Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio1 found a few suspects,
-including the combination of the popular logistic sigmoid activation function and the weight initialization
-technique that was most popular at the time, namely random initialization using a normal distribution with
-a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this
-initialization scheme, the variance of the outputs of each layer is much greater than the variance of its
-inputs. Going forward in the network, the variance keeps increasing after each layer until the activation
-function saturates at the top layers. This is actually made worse by the fact that the logistic function has a
-mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the
-logistic function in deep networks).
+A paper titled Understanding the Difficulty of Training Deep
+Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio found that
+the problems with the popular logistic
+sigmoid activation function and the weight initialization technique
+that was most popular at the time, namely random initialization using
+a normal distribution with a mean of 0 and a standard deviation of
+1.
+
+
          
+They showed that with this activation function and this
+initialization scheme, the variance of the outputs of each layer is
+much greater than the variance of its inputs. Going forward in the
+network, the variance keeps increasing after each layer until the
+activation function saturates at the top layers. This is actually made
+worse by the fact that the logistic function has a mean of 0.5, not 0
+(the hyperbolic tangent function has a mean of 0 and behaves slightly
+better than the logistic function in deep networks).
 
 
          
 









          
 
-
          The derivative of the Logistic funtion 
          
+
          The derivative of the Logistic funtion 
          
 
 
          
 Looking at the logistic activation function, when inputs become large
-(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when
-backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what
-little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so
-there is really nothing left for the lower layers.
+(negative or positive), the function saturates at 0 or 1, with a
+derivative extremely close to 0. Thus when backpropagation kicks in,
+it has virtually no gradient to propagate back through the network,
+and what little gradient exists keeps getting diluted as
+backpropagation progresses down through the top layers, so there is
+really nothing left for the lower layers.
 
 
          
-In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the
-signal to flow properly in both directions: in the forward direction when making predictions, and in the
-reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it
-to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the
-outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have
-equal variance before and after flowing through a layer in the reverse direction (please check out the
-paper if you are interested in the mathematical details).
+In their paper, Glorot and Bengio propose a way to significantly
+alleviate this problem. We need the signal to flow properly in both
+directions: in the forward direction when making predictions, and in
+the reverse direction when backpropagating gradients. We don’t want
+the signal to die out, nor do we want it to explode and saturate. For
+the signal to flow properly, the authors argue that we need the
+variance of the outputs of each layer to be equal to the variance of
+its inputs, and we also need the gradients to have equal variance
+before and after flowing through a layer in the reverse direction.
 
 
          
-One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients
-problems were in part due to a poor choice of activation function. Until then most people had assumed
-that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they
-must be an excellent choice. But it turns out that other activation functions behave much better in deep
-neural networks, in particular the ReLU activation function, mostly because it does not saturate for
-positive values (and also because it is quite fast to compute).
+One of the insights in the 2010 paper by Glorot and Bengio was that
+the vanishing/exploding gradients problems were in part due to a poor
+choice of activation function. Until then most people had assumed that
+if Nature had chosen to use roughly sigmoid activation functions in
+biological neurons, they must be an excellent choice. But it turns out
+that other activation functions behave much better in deep neural
+networks, in particular the ReLU activation function, mostly because
+it does not saturate for positive values (and also because it is quite
+fast to compute).
 
 
          
 









          
 
-
          The RELU function family 
          
+
          The RELU function family 
          
 
 
          
 The ReLU activation function suffers from a problem known as the dying
-ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.
+ReLUs: during training, some neurons effectively die, meaning they
+stop outputting anything other than 0.
 
 
          
-In some cases, you may find that half of your network’s neurons are dead, especially if you used a large
-learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s
-inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life
-since the gradient of the ReLU function is 0 when its input is negative.
+In some cases, you may find that half of your network’s neurons are
+dead, especially if you used a large learning rate. During training,
+if a neuron’s weights get updated such that the weighted sum of the
+neuron’s inputs is negative, it will start outputting 0. When this
+happen, the neuron is unlikely to come back to life since the gradient
+of the ReLU function is 0 when its input is negative.
 
 
          
-To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function
+To solve this problem, nowadays practitioners use a  variant of the ReLU
+function, such as the leaky ReLU discussed above or the so-called
+exponential linear unit (ELU) function
+
 $$
 ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
 $$
 
 
          
-So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,
-in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than \( \tanh \) which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you
-may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default \( \alpha \) of
-\( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have spare time and computing power, you can use
-cross-validation or bootstrap to evaluate other activation functions.
-huge training set.
+









          
+
+
          Which activation function should we use? 
          
+
+
          
+In general it seems that the ELU activation function is better than
+the leaky ReLU function (and its variants), which is better than
+ReLU. ReLU performs better than \( \tanh \) which in turn performs better
+than the logistic function.
+
+
          
+If runtime
+performance is an issue, then you may opt for the  leaky ReLU function  over the 
+ELU function If you don’t
+want to tweak yet another hyperparameter, you may just use the default
+\( \alpha \) of \( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have
+spare time and computing power, you can use cross-validation or
+bootstrap to evaluate other activation functions.
 
 
          
 
 
-
          A top-down perspective on Neural networks 
          
+
          A top-down perspective on Neural networks 
          
 
 
          
 The first thing we would like to do is divide the data into two or three
@@ -2765,8 +2857,10 @@ cardinal sin in ML. Then:
 
      
 
 If the validation and test sets are drawn from the same distributions,
-then good performance on the validation set should lead to similarly
-good performance on the test set. 
+then a good performance on the validation set should lead to similarly
+good performance on the test set.
+
+
      
 However, sometimes
 the training data and test data differ in subtle ways because, for
 example, they are collected using slightly different methods, or
@@ -2785,7 +2879,7 @@ supervised learning.
 
      
 









      
 
-
      Limitations of supervised learning with deep networks 
      
+
      Limitations of supervised learning with deep networks 
      
 
 
      
 Like all statistical methods, supervised learning using neural
diff --git a/doc/pub/NeuralNet/html/NeuralNet.html b/doc/pub/NeuralNet/html/NeuralNet.html
index e468079e7..b0a0523e3 100644
--- a/doc/pub/NeuralNet/html/NeuralNet.html
+++ b/doc/pub/NeuralNet/html/NeuralNet.html
@@ -116,53 +116,55 @@ div { text-align: justify; text-justify: inter-word; }
                None,
                '___sec30'),
               ('Defining the cost function', 2, None, '___sec31'),
+              ('Example: binary classification problem', 2, None, '___sec32'),
               ('Developing a code for doing neural networks with back '
                'propagation',
                2,
                None,
-               '___sec32'),
-              ('Collect and pre-process data', 2, None, '___sec33'),
-              ('Train and test datasets', 2, None, '___sec34'),
-              ('Define model and architecture', 2, None, '___sec35'),
-              ('Layers', 2, None, '___sec36'),
-              ('Weights and biases', 2, None, '___sec37'),
-              ('Feed-forward pass', 2, None, '___sec38'),
-              ('Matrix multiplications', 2, None, '___sec39'),
-              ('Choose cost function and optimizer', 2, None, '___sec40'),
-              ('Optimizing the cost function', 2, None, '___sec41'),
-              ('Regularization', 2, None, '___sec42'),
-              ('Matrix  multiplication', 2, None, '___sec43'),
-              ('Improving performance', 2, None, '___sec44'),
-              ('Full object-oriented implementation', 2, None, '___sec45'),
-              ('Evaluate model performance on test data', 2, None, '___sec46'),
-              ('Adjust hyperparameters', 2, None, '___sec47'),
-              ('Visualization', 2, None, '___sec48'),
-              ('scikit-learn implementation', 2, None, '___sec49'),
-              ('Visualization', 2, None, '___sec50'),
+               '___sec33'),
+              ('Collect and pre-process data', 2, None, '___sec34'),
+              ('Train and test datasets', 2, None, '___sec35'),
+              ('Define model and architecture', 2, None, '___sec36'),
+              ('Layers', 2, None, '___sec37'),
+              ('Weights and biases', 2, None, '___sec38'),
+              ('Feed-forward pass', 2, None, '___sec39'),
+              ('Matrix multiplications', 2, None, '___sec40'),
+              ('Choose cost function and optimizer', 2, None, '___sec41'),
+              ('Optimizing the cost function', 2, None, '___sec42'),
+              ('Regularization', 2, None, '___sec43'),
+              ('Matrix  multiplication', 2, None, '___sec44'),
+              ('Improving performance', 2, None, '___sec45'),
+              ('Full object-oriented implementation', 2, None, '___sec46'),
+              ('Evaluate model performance on test data', 2, None, '___sec47'),
+              ('Adjust hyperparameters', 2, None, '___sec48'),
+              ('Visualization', 2, None, '___sec49'),
+              ('scikit-learn implementation', 2, None, '___sec50'),
+              ('Visualization', 2, None, '___sec51'),
               ('Building neural networks in Tensorflow and Keras',
                2,
                None,
-               '___sec51'),
-              ('Tensorflow', 2, None, '___sec52'),
-              ('Collect and pre-process data', 2, None, '___sec53'),
-              ('Using TensorFlow backend', 2, None, '___sec54'),
-              ('Optimizing and using gradient descent', 2, None, '___sec55'),
-              ('Using Keras', 2, None, '___sec56'),
-              ('Which activation function should I use?', 2, None, '___sec57'),
+               '___sec52'),
+              ('Tensorflow', 2, None, '___sec53'),
+              ('Collect and pre-process data', 2, None, '___sec54'),
+              ('Using TensorFlow backend', 2, None, '___sec55'),
+              ('Optimizing and using gradient descent', 2, None, '___sec56'),
+              ('Using Keras', 2, None, '___sec57'),
+              ('Which activation function should I use?', 2, None, '___sec58'),
               ('Is the Logistic activation function (Sigmoid)  our choice?',
                2,
                None,
-               '___sec58'),
-              ('The derivative of the Logistic funtion', 2, None, '___sec59'),
-              ('The RELU function family', 2, None, '___sec60'),
+               '___sec59'),
+              ('The derivative of the Logistic funtion', 2, None, '___sec60'),
+              ('The RELU function family', 2, None, '___sec61'),
+              ('Which activation function should we use?', 2, None, '___sec62'),
               ('A top-down perspective on Neural networks',
                2,
                None,
-               '___sec61'),
+               '___sec63'),
               ('Limitations of supervised learning with deep networks',
                2,
                None,
-               '___sec62')]}
+               '___sec64')]}
 end of tocinfo -->
 
 
@@ -204,7 +206,7 @@ MathJax.Hub.Config({
 
      [2] Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University
      
 
      
 
      
-
      Oct 12, 2018
       
+
      Oct 16, 2018
       
 
      
 
      
 









      
@@ -1089,7 +1091,7 @@ $$
 z_j^{l+1} = \sum_{i=1}^{M_{l}}w_{ij}^{l+1}a_j^{l}+b_j^{l+1},
 $$
 
-we obtain
+with \( M_l \) being the number of nodes in layer \( l \), we obtain
 $$
 \delta_j^l =\sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l),
 $$
@@ -1145,7 +1147,7 @@ $$
 
      
 Then we compute the back propagate error for each \( l=L-1,L-2,\dots,2 \) as
 $$
-\delta_j^l =\sum_k \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
+\delta_j^l = \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
 $$
 
    
 
@@ -1154,14 +1156,14 @@ $$
 
    
 
 
    
-Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,dots,2 \) and update the weights and biases according to the rules
+Finally, we update the weights and the biases using gradient descent for each \( l=L-1,L-2,\dots,2 \) and update the weights and biases according to the rules
 $$
 w_{jk}^l\leftarrow  = w_{jk}^l- \eta \delta_j^la_k^{l-1},
 $$
 
 
 $$
-b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^L},
+b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^l}=b_j^l-\eta \delta_j^l,
 $$
 
    
 
@@ -1260,19 +1262,66 @@ $$
 Again we take the negative log-likelihood to define our cost function:  
 
 $$
-\mathcal{C}(\hat{\theta}) = - \ln P(\mathcal{D} \mid \hat{\theta}).
+\mathcal{C}(\hat{\theta}) = - \log{P(\mathcal{D} \mid \hat{\theta})}.
 $$
 
 See the logistic regression lectures for a full definition of the cost function.
 
 
    
 The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before!
-We leave it as an exercise in project 2 to derive these equations.
+
+
    
+









    
+
+
    Example: binary classification problem 
    
+
+
    
+As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters \( \beta \) as
+$$
+\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right),
+$$
+
+where we had defined the logistic (sigmoid) function
+$$
+p(y_i =1\vert x_i,\hat{\beta})=\frac{\exp{(\beta_0+\beta_1 x_i)}}{1+\exp{(\beta_0+\beta_1 x_i)}},
+$$
+
+and
+$$
+p(y_i =0\vert x_i,\hat{\beta})=1-p(y_i =1\vert x_i,\hat{\beta}).
+$$
+
+The parameters \( \hat{\beta} \) were defined using a minimization method like gradient descent or Newton-Raphson's method.
+
+
    
+Now we replace \( x_i \) with the activation \( z_i^l \) for a given layer \( l \) and the outputs as \( y_i=a_i^l=f(z_i^l) \), with \( z_i^l \) now being a function of the weights \( w_{ij}^l \) and biases \( b_i^l \). 
+We have then
+$$
+a_i^l = y_i = \frac{\exp{(z_i^l)}}{1+\exp{(z_i^l)}},
+$$
+
+with 
+$$
+z_i^l = \sum_{j}w_{ij}^l a_j^{l-1}+b_i^l,
+$$
+
+where the superscript \( l-1 \) indicates that these are the outputs from layer \( l-1 \).
+Our cost function at the final layer \( l=L \) is now
+$$
+\mathcal{C}(\hat{W}) = - \sum_{i=1}^n \left(t_i\log{a_i^L}+(i-t_i)\log{(1-a_i^L)}\right),
+$$
+
+where we have defined the targets \( t_i \). The derivatives of the cost function with respect to the output \( a_i^L \) are then easily calculated and we get
+$$
+\frac{\partial \mathcal{C}(\hat{W})}{\partial a_i^L} = \frac{a_i^L-t_i}{a_i^L(1-a_i^L)}. 
+$$
+
+In case we use another activation function than the logistic one, we need to evaluate other derivatives.
 
 
    
 
 
-
    Developing a code for doing neural networks with back propagation 
    
+
    Developing a code for doing neural networks with back propagation 
    
 
 
    
 One can identify a set of key steps when using neural networks to solve supervised learning problems:  
@@ -1288,7 +1337,7 @@ One can identify a set of key steps when using neural networks to solve supervis
 
 









    
 
-
    Collect and pre-process data 
    
+
    Collect and pre-process data 
    
 
 
    
 Here we will be using the MNIST dataset, which is readily available through the scikit-learn
@@ -1383,7 +1432,7 @@ plt.show()
 
    
 









    
 
-
    Train and test datasets 
    
+
    Train and test datasets 
    
 
 
    
 Performing analysis before partitioning the dataset is a major error, that can lead to incorrect conclusions.
@@ -1432,7 +1481,7 @@ X_train, X_test, Y_train, Y_test = train_tes
 
    
 









    
 
-
    Define model and architecture 
    
+
    Define model and architecture 
    
 
 
    
 Our simple feed-forward neural network will consist of an input layer, a single hidden layer and an output layer. The activation \( y \) of each neuron is a weighted sum of inputs, passed through an activation function. In case of the simple perceptron model we have 
@@ -1473,12 +1522,12 @@ We will be using the sigmoid function \( \sigma(x) \):
 $$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
 
 
    
-which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.
+which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.
 
 
    
 
 
-
    Layers 
    
+
    Layers 
    
 
 
      
 
    •  Input
      •  
@@ -1523,7 +1572,7 @@ weights to the output layer.
 
      
 
 
-
      Weights and biases 
      
+
      Weights and biases 
      
 
 
      
 Typically weights are initialized with small values distributed around zero, drawn from a uniform
@@ -1559,7 +1608,7 @@ output_bias = np
 









      
 
-
      Feed-forward pass 
      
+
      Feed-forward pass 
      
 
 
      
 Denote \( F \) the number of features, \( H \) the number of hidden neurons and \( C \) the number of categories.  
@@ -1586,7 +1635,7 @@ $$ a_{j}^{L} = \frac{\exp{(z_j^{L})}}
 
      
 
 
-
      Matrix multiplications 
      
+
      Matrix multiplications 
      
 
 
      
 Since our data has the dimensions \( X = (n_{inputs}, n_{features}) \) and our weights to the hidden
@@ -1661,7 +1710,7 @@ predictions = predict(X_train)
 
      
 









      
 
-
      Choose cost function and optimizer 
      
+
      Choose cost function and optimizer 
      
 
 
      
 To measure how well our neural network is doing we need to introduce a cost function.  
@@ -1692,7 +1741,7 @@ you got the correct label. The probability of category \( c \) is given by the s
 
      
 









      
 
-
      Optimizing the cost function 
      
+
      Optimizing the cost function 
      
 
 
      
 The network is trained by finding the weights and biases that minimize the cost function. One of the most widely used classes of methods is gradient descent and its generalizations. The idea behind gradient descent
@@ -1727,9 +1776,12 @@ This has two important benefits:
 
    •  It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
      •   
 
 
+The various optmization  methods, with codes and algorithms,  are discussed in our lectures on Gradient descent approaches.
+
+
      
 
 
-
      Regularization 
      
+
      Regularization 
      
 
 
      
 It is common to add an extra term to the cost function, proportional
@@ -1762,13 +1814,13 @@ calculate the gradient efficently.
 
      
 









      
 
-
      Matrix  multiplication 
      
+
      Matrix  multiplication 
      
 
 
      
 To more efficently train our network these equations are implemented using matrix operations.  
-The error in the output layer is calculated simply as  
+The error in the output layer is calculated simply as, with \( \hat{t} \) being our targets,  
 
-$$ \delta_L = \hat{y} - y = (n_{inputs}, n_{categories}) .$$
+$$ \delta_L = \hat{t} - \hat{y} = (n_{inputs}, n_{categories}) .$$
 
 
      
 The gradient for the output weights is calculated as  
@@ -1805,9 +1857,6 @@ $$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
 
 
 
      # to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our model
 from sklearn.metrics import accuracy_score
 
 # one-hot in numpy
@@ -1879,7 +1928,7 @@ lmbd = 0.01
 









      
 
-
      Improving performance 
      
+
      Improving performance 
      
 
 
      
 As we can see the network does not seem to be learning at all. It seems to be just guessing the label for each image.  
@@ -1899,7 +1948,7 @@ Andrew Ng goes through some of these considerations in this Full object-oriented implementation 
+
      Full object-oriented implementation 
      
 
 
      
 It is very natural to think of the network as an object, with specific instances of the network
@@ -2012,7 +2061,7 @@ being realizations of this object with different hyperparameters. An implementat
 
      
 









      
 
-
      Evaluate model performance on test data 
      
+
      Evaluate model performance on test data 
      
 
 
      
 To measure the performance of our network we evaluate how well it does it data it has never seen before, i.e. the test data.  
@@ -2047,7 +2096,7 @@ test_predict = dnnAdjust hyperparameters 
+
      Adjust hyperparameters 
      
 
 
      
 We now perform a grid search to find the optimal hyperparameters for the network.  
@@ -2080,7 +2129,7 @@ DNN_numpy = np.
 
      
 









      
 
-
      Visualization 
      
+
      Visualization 
      
 
 
      
 
@@ -2122,7 +2171,7 @@ plt.show()
 
      
 









      
 
-
      scikit-learn implementation 
      
+
      scikit-learn implementation 
      
 
 
      
 scikit-learn focuses more
@@ -2161,7 +2210,7 @@ DNN_scikit = np
 
      
 









      
 
-
      Visualization 
      
+
      Visualization 
      
 
      
 
 
@@ -2203,7 +2252,7 @@ plt.show()
 
      
 









      
 
-
      Building neural networks in Tensorflow and Keras 
      
+
      Building neural networks in Tensorflow and Keras 
      
 
 
      
 Now we want  to build on the experience gained from our neural network implementation in NumPy and scikit-learn
@@ -2218,7 +2267,7 @@ NumPy arrays.
 
      
 









      
 
-
      Tensorflow 
      
+
      Tensorflow 
      
 
 
      
 Tensorflow is an open source library machine learning library
@@ -2263,7 +2312,7 @@ and/or if you use anaconda, just write (or install from the graphical use
 
      
 









      
 
-
      Collect and pre-process data 
      
+
      Collect and pre-process data 
      
 
 
      
 
@@ -2329,7 +2378,7 @@ X_train, X_test, Y_train, Y_test = train_tes
 
      
 









      
 
-
      Using TensorFlow backend 
      
+
      Using TensorFlow backend 
      
 
 
        
 
      1.  Define model and architecture
        2. 
@@ -2474,7 +2523,7 @@ X_train, X_test, Y_train, Y_test = train_tes
 
        
 









        
 
-
        Optimizing and using gradient descent 
        
+
        Optimizing and using gradient descent 
        
 
 
        
 
@@ -2552,7 +2601,7 @@ writer.add_graph(tfUsing Keras 
+
        Using Keras 
        
 
 
        
 Keras is a high level neural network
@@ -2651,106 +2700,149 @@ plt.show()
 
        
 
 
-
        Which activation function should I use? 
        
+
        Which activation function should I use? 
        
 
 
        
-Backpropagation algorithm works by going from the output layer to the
-input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the
-cost function with regards to each parameter in the network, it uses these gradients to update each
-parameter with a Gradient Descent step.
+The Back propagation algorithm we derived above works by going from
+the output layer to the input layer, propagating the error gradient on
+the way. Once the algorithm has computed the gradient of the cost
+function with regards to each parameter in the network, it uses these
+gradients to update each parameter with a Gradient Descent (GD) step.
 
 
        
-Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower
-layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually
-unchanged, and training never converges to a good solution. This is called the vanishing gradients
-problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many
-layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients
-problem, which is mostly encountered in recurrent neural networks. More generally,
-deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds
+Unfortunately for us, the gradients often get smaller and smaller as the
+algorithm progresses down to the first hidden layers. As a result, the
+GD update leaves the lower layer connection weights
+virtually unchanged, and training never converges to a good
+solution. This is known in the literature as 
+the vanishing gradients problem.
+
+
        
+In other cases, the opposite can happen, namely the the gradients can grow bigger and
+bigger. The result is that many of the layers get large updates of the 
+weights the
+algorithm diverges. This is the exploding gradients problem, which is
+mostly encountered in recurrent neural networks. More generally, deep
+neural networks suffer from unstable gradients, different layers may
+learn at widely different speeds
 
 
        
 
 
-
        Is the Logistic activation function (Sigmoid)  our choice? 
        
+
        Is the Logistic activation function (Sigmoid)  our choice? 
        
 
 
        
-Although this unfortunate behavior has been empirically observed for quite a while (it was one of the
-reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that
+Although this unfortunate behavior has been empirically observed for
+quite a while (it was one of the reasons why deep neural networks were
+mostly abandoned for a long time), it is only around 2010 that
 significant progress was made in understanding it.
 
 
        
-A paper titled Understanding the Difficulty of Training Deep Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio1 found a few suspects,
-including the combination of the popular logistic sigmoid activation function and the weight initialization
-technique that was most popular at the time, namely random initialization using a normal distribution with
-a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this
-initialization scheme, the variance of the outputs of each layer is much greater than the variance of its
-inputs. Going forward in the network, the variance keeps increasing after each layer until the activation
-function saturates at the top layers. This is actually made worse by the fact that the logistic function has a
-mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the
-logistic function in deep networks).
+A paper titled Understanding the Difficulty of Training Deep
+Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio found that
+the problems with the popular logistic
+sigmoid activation function and the weight initialization technique
+that was most popular at the time, namely random initialization using
+a normal distribution with a mean of 0 and a standard deviation of
+1.
+
+
        
+They showed that with this activation function and this
+initialization scheme, the variance of the outputs of each layer is
+much greater than the variance of its inputs. Going forward in the
+network, the variance keeps increasing after each layer until the
+activation function saturates at the top layers. This is actually made
+worse by the fact that the logistic function has a mean of 0.5, not 0
+(the hyperbolic tangent function has a mean of 0 and behaves slightly
+better than the logistic function in deep networks).
 
 
        
 









        
 
-
        The derivative of the Logistic funtion 
        
+
        The derivative of the Logistic funtion 
        
 
 
        
 Looking at the logistic activation function, when inputs become large
-(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when
-backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what
-little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so
-there is really nothing left for the lower layers.
+(negative or positive), the function saturates at 0 or 1, with a
+derivative extremely close to 0. Thus when backpropagation kicks in,
+it has virtually no gradient to propagate back through the network,
+and what little gradient exists keeps getting diluted as
+backpropagation progresses down through the top layers, so there is
+really nothing left for the lower layers.
 
 
        
-In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the
-signal to flow properly in both directions: in the forward direction when making predictions, and in the
-reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it
-to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the
-outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have
-equal variance before and after flowing through a layer in the reverse direction (please check out the
-paper if you are interested in the mathematical details).
+In their paper, Glorot and Bengio propose a way to significantly
+alleviate this problem. We need the signal to flow properly in both
+directions: in the forward direction when making predictions, and in
+the reverse direction when backpropagating gradients. We don’t want
+the signal to die out, nor do we want it to explode and saturate. For
+the signal to flow properly, the authors argue that we need the
+variance of the outputs of each layer to be equal to the variance of
+its inputs, and we also need the gradients to have equal variance
+before and after flowing through a layer in the reverse direction.
 
 
        
-One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients
-problems were in part due to a poor choice of activation function. Until then most people had assumed
-that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they
-must be an excellent choice. But it turns out that other activation functions behave much better in deep
-neural networks, in particular the ReLU activation function, mostly because it does not saturate for
-positive values (and also because it is quite fast to compute).
+One of the insights in the 2010 paper by Glorot and Bengio was that
+the vanishing/exploding gradients problems were in part due to a poor
+choice of activation function. Until then most people had assumed that
+if Nature had chosen to use roughly sigmoid activation functions in
+biological neurons, they must be an excellent choice. But it turns out
+that other activation functions behave much better in deep neural
+networks, in particular the ReLU activation function, mostly because
+it does not saturate for positive values (and also because it is quite
+fast to compute).
 
 
        
 









        
 
-
        The RELU function family 
        
+
        The RELU function family 
        
 
 
        
 The ReLU activation function suffers from a problem known as the dying
-ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.
+ReLUs: during training, some neurons effectively die, meaning they
+stop outputting anything other than 0.
 
 
        
-In some cases, you may find that half of your network’s neurons are dead, especially if you used a large
-learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s
-inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life
-since the gradient of the ReLU function is 0 when its input is negative.
+In some cases, you may find that half of your network’s neurons are
+dead, especially if you used a large learning rate. During training,
+if a neuron’s weights get updated such that the weighted sum of the
+neuron’s inputs is negative, it will start outputting 0. When this
+happen, the neuron is unlikely to come back to life since the gradient
+of the ReLU function is 0 when its input is negative.
 
 
        
-To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function
+To solve this problem, nowadays practitioners use a  variant of the ReLU
+function, such as the leaky ReLU discussed above or the so-called
+exponential linear unit (ELU) function
+
 $$
 ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
 $$
 
 
        
-So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,
-in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than \( \tanh \) which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you
-may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default \( \alpha \) of
-\( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have spare time and computing power, you can use
-cross-validation or bootstrap to evaluate other activation functions.
-huge training set.
+









        
+
+
        Which activation function should we use? 
        
+
+
        
+In general it seems that the ELU activation function is better than
+the leaky ReLU function (and its variants), which is better than
+ReLU. ReLU performs better than \( \tanh \) which in turn performs better
+than the logistic function.
+
+
        
+If runtime
+performance is an issue, then you may opt for the  leaky ReLU function  over the 
+ELU function If you don’t
+want to tweak yet another hyperparameter, you may just use the default
+\( \alpha \) of \( 0.01 \) for the leaky ReLU, and \( 1 \) for ELU. If you have
+spare time and computing power, you can use cross-validation or
+bootstrap to evaluate other activation functions.
 
 
        
 
 
-
        A top-down perspective on Neural networks 
        
+
        A top-down perspective on Neural networks 
        
 
 
        
 The first thing we would like to do is divide the data into two or three
@@ -2770,8 +2862,10 @@ cardinal sin in ML. Then:
 
    
 
 If the validation and test sets are drawn from the same distributions,
-then good performance on the validation set should lead to similarly
-good performance on the test set. 
+then a good performance on the validation set should lead to similarly
+good performance on the test set.
+
+
    
 However, sometimes
 the training data and test data differ in subtle ways because, for
 example, they are collected using slightly different methods, or
@@ -2790,7 +2884,7 @@ supervised learning.
 
    
 









    
 
-
    Limitations of supervised learning with deep networks 
    
+
    Limitations of supervised learning with deep networks 
    
 
 
    
 Like all statistical methods, supervised learning using neural
diff --git a/doc/pub/NeuralNet/ipynb/NeuralNet.ipynb b/doc/pub/NeuralNet/ipynb/NeuralNet.ipynb
index c184e566a..c0c8483f6 100644
--- a/doc/pub/NeuralNet/ipynb/NeuralNet.ipynb
+++ b/doc/pub/NeuralNet/ipynb/NeuralNet.ipynb
@@ -10,7 +10,7 @@
     "  \n",
     "**Morten Hjorth-Jensen**, Department of Physics, University of Oslo and Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University\n",
     "\n",
-    "Date: **Oct 12, 2018**\n",
+    "Date: **Oct 16, 2018**\n",
     "\n",
     "Copyright 1999-2018, Morten Hjorth-Jensen. Released under CC Attribution-NonCommercial 4.0 license\n",
     "\n",
@@ -596,50 +596,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 13,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    }
-   ],
+   "execution_count": 1,
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "%matplotlib inline\n",
     "\n",
@@ -1248,7 +1209,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "we obtain"
+    "with $M_l$ being the number of nodes in layer $l$, we obtain"
    ]
   },
   {
@@ -1311,7 +1272,7 @@
    "metadata": {},
    "source": [
     "$$\n",
-    "\\delta_j^l =\\sum_k \\sum_k \\delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).\n",
+    "\\delta_j^l = \\sum_k \\delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).\n",
     "$$"
    ]
   },
@@ -1319,7 +1280,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "Finally, we update the weights and the biases using gradient descent for each $l=L-1,L-2,dots,2$ and update the weights and biases according to the rules"
+    "Finally, we update the weights and the biases using gradient descent for each $l=L-1,L-2,\\dots,2$ and update the weights and biases according to the rules"
    ]
   },
   {
@@ -1336,7 +1297,7 @@
    "metadata": {},
    "source": [
     "$$\n",
-    "b_j^l \\leftarrow b_j^l-\\eta \\frac{\\partial {\\cal C}}{\\partial b_j^L},\n",
+    "b_j^l \\leftarrow b_j^l-\\eta \\frac{\\partial {\\cal C}}{\\partial b_j^l}=b_j^l-\\eta \\delta_j^l,\n",
     "$$"
    ]
   },
@@ -1482,7 +1443,7 @@
    "metadata": {},
    "source": [
     "$$\n",
-    "\\mathcal{C}(\\hat{\\theta}) = - \\ln P(\\mathcal{D} \\mid \\hat{\\theta}).\n",
+    "\\mathcal{C}(\\hat{\\theta}) = - \\log{P(\\mathcal{D} \\mid \\hat{\\theta})}.\n",
     "$$"
    ]
   },
@@ -1493,7 +1454,126 @@
     "See the logistic regression lectures for a full definition of the cost function.\n",
     "\n",
     "The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before!\n",
-    "We leave it as an exercise in project 2 to derive these equations.\n",
+    "\n",
+    "## Example: binary classification problem\n",
+    "\n",
+    "As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters $\\beta$ as"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "\\mathcal{C}(\\hat{\\beta}) = - \\sum_{i=1}^n \\left(y_i\\log{p(y_i \\vert x_i,\\hat{\\beta})}+(i-y_i)\\log{1-p(y_i \\vert x_i,\\hat{\\beta})}\\right),\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "where we had defined the logistic (sigmoid) function"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "p(y_i =1\\vert x_i,\\hat{\\beta})=\\frac{\\exp{(\\beta_0+\\beta_1 x_i)}}{1+\\exp{(\\beta_0+\\beta_1 x_i)}},\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "and"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "p(y_i =0\\vert x_i,\\hat{\\beta})=1-p(y_i =1\\vert x_i,\\hat{\\beta}).\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "The parameters $\\hat{\\beta}$ were defined using a minimization method like gradient descent or Newton-Raphson's method. \n",
+    "\n",
+    "Now we replace $x_i$ with the activation $z_i^l$ for a given layer $l$ and the outputs as $y_i=a_i^l=f(z_i^l)$, with $z_i^l$ now being a function of the weights $w_{ij}^l$ and biases $b_i^l$. \n",
+    "We have then"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "a_i^l = y_i = \\frac{\\exp{(z_i^l)}}{1+\\exp{(z_i^l)}},\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "with"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "z_i^l = \\sum_{j}w_{ij}^l a_j^{l-1}+b_i^l,\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "where the superscript $l-1$ indicates that these are the outputs from layer $l-1$.\n",
+    "Our cost function at the final layer $l=L$ is now"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "\\mathcal{C}(\\hat{W}) = - \\sum_{i=1}^n \\left(t_i\\log{a_i^L}+(i-t_i)\\log{(1-a_i^L)}\\right),\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "where we have defined the targets $t_i$. The derivatives of the cost function with respect to the output $a_i^L$ are then easily calculated and we get"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "$$\n",
+    "\\frac{\\partial \\mathcal{C}(\\hat{W})}{\\partial a_i^L} = \\frac{a_i^L-t_i}{a_i^L(1-a_i^L)}.\n",
+    "$$"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In case we use another activation function than the logistic one, we need to evaluate other derivatives. \n",
     "\n",
     "\n",
     "\n",
@@ -1558,29 +1638,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "inputs = (n_inputs, pixel_width, pixel_height) = (1797, 8, 8)\n",
-      "labels = (n_inputs) = (1797,)\n",
-      "X = (n_inputs, n_features) = (1797, 64)\n"
-     ]
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    }
-   ],
+   "execution_count": 2,
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "# import necessary packages\n",
     "import numpy as np\n",
@@ -1645,18 +1707,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Number of training images: 1437\n",
-      "Number of test images: 360\n"
-     ]
-    }
-   ],
+   "execution_count": 3,
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "from sklearn.model_selection import train_test_split\n",
     "\n",
@@ -1725,7 +1780,7 @@
     "\n",
     "$$ f(x) = \\sigma(x) = \\frac{1}{1 + e^{-x}} ,$$\n",
     "\n",
-    "which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.\n",
+    "which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.\n",
     "\n",
     "\n",
     "## Layers\n",
@@ -1777,7 +1832,9 @@
   {
    "cell_type": "code",
    "execution_count": 4,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "# building our neural network\n",
@@ -1854,25 +1911,10 @@
   {
    "cell_type": "code",
    "execution_count": 5,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "probabilities = (n_inputs, n_categories) = (1437, 10)\n",
-      "probability that image 0 is in category 0,1,2,...,9 = \n",
-      "[5.41511965e-04 2.17174962e-03 8.84355903e-03 1.44970586e-03\n",
-      " 1.10378326e-04 5.08318298e-09 2.03256632e-04 1.92507116e-03\n",
-      " 9.84443254e-01 3.11507992e-04]\n",
-      "probabilities sum up to: 1.0\n",
-      "\n",
-      "predictions = (n_inputs) = (1437,)\n",
-      "prediction for image 0: 8\n",
-      "correct label for image 0: 6\n"
-     ]
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "# setup the feed-forward pass, subscript h = hidden layer\n",
     "\n",
@@ -1969,6 +2011,8 @@
     "\n",
     "2. It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.  \n",
     "\n",
+    "The various optmization  methods, with codes and algorithms,  are discussed in our lectures on [Gradient descent approaches](https://compphysics.github.io/MachineLearning/doc/pub/Splines/html/Splines-bs.html).\n",
+    "\n",
     "\n",
     "## Regularization\n",
     "\n",
@@ -2001,9 +2045,9 @@
     "## Matrix  multiplication\n",
     "\n",
     "To more efficently train our network these equations are implemented using matrix operations.  \n",
-    "The error in the output layer is calculated simply as  \n",
+    "The error in the output layer is calculated simply as, with $\\hat{t}$ being our targets,  \n",
     "\n",
-    "$$ \\delta_L = \\hat{y} - y = (n_{inputs}, n_{categories}) .$$  \n",
+    "$$ \\delta_L = \\hat{t} - \\hat{y} = (n_{inputs}, n_{categories}) .$$  \n",
     "\n",
     "The gradient for the output weights is calculated as  \n",
     "\n",
@@ -2034,36 +2078,12 @@
   {
    "cell_type": "code",
    "execution_count": 6,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Old accuracy on training data: 0.1440501043841336\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/ipykernel_launcher.py:4: RuntimeWarning: overflow encountered in exp\n",
-      "  after removing the cwd from sys.path.\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "New accuracy on training data: 0.09672929714683368\n"
-     ]
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "# to categorical turns our integer vector into a onehot representation\n",
-    "#from keras.utils import to_categorical\n",
-    "\n",
-    "# calculate the accuracy score of our model\n",
     "from sklearn.metrics import accuracy_score\n",
     "\n",
     "# one-hot in numpy\n",
@@ -2159,7 +2179,9 @@
   {
    "cell_type": "code",
    "execution_count": 7,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "class NeuralNetwork:\n",
@@ -2282,16 +2304,10 @@
   {
    "cell_type": "code",
    "execution_count": 8,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Accuracy score on test set:  0.9416666666666667\n"
-     ]
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "epochs = 100\n",
     "batch_size = 100\n",
@@ -2324,239 +2340,10 @@
   {
    "cell_type": "code",
    "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.11666666666666667\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.20833333333333334\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.12222222222222222\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.14722222222222223\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.17777777777777778\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.16111111111111112\n",
-      "\n",
-      "Learning rate  =  1e-05\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.20277777777777778\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.5305555555555556\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.5944444444444444\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.5888888888888889\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.6111111111111112\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.5222222222222223\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.5555555555555556\n",
-      "\n",
-      "Learning rate  =  0.0001\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.8055555555555556\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.85\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.85\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.875\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.8666666666666667\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.8638888888888889\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.9555555555555556\n",
-      "\n",
-      "Learning rate  =  0.001\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.925\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.9333333333333333\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.9472222222222222\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.9416666666666667\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.9194444444444444\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.9416666666666667\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.8222222222222222\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.23333333333333334\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/ipykernel_launcher.py:4: RuntimeWarning: overflow encountered in exp\n",
-      "  after removing the cwd from sys.path.\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.1\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.10555555555555556\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.125\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.08888888888888889\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.125\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.09166666666666666\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.09166666666666666\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/ipykernel_launcher.py:44: RuntimeWarning: overflow encountered in exp\n",
-      "/usr/local/lib/python3.7/site-packages/ipykernel_launcher.py:45: RuntimeWarning: invalid value encountered in true_divide\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1.0\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.07777777777777778\n",
-      "\n"
-     ]
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "eta_vals = np.logspace(-5, 1, 7)\n",
     "lmbd_vals = np.logspace(-5, 1, 7)\n",
@@ -2590,37 +2377,10 @@
   {
    "cell_type": "code",
    "execution_count": 10,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/ipykernel_launcher.py:4: RuntimeWarning: overflow encountered in exp\n",
-      "  after removing the cwd from sys.path.\n"
-     ]
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "data": {
-      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAkgAAAJgCAYAAABx4Zz0AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4wLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvqOYd8AAAIABJREFUeJzs3XucjHX/x/HXzJ5YdlfLHpxVDiuHnEM5JzlEKjdRDpU7yiGV6BaldHL2KzkkkVOH21m4iUJZh1BEKIeE3bVr2bPd2Znr98do1jVDSnbW1Pv5eMyD65pr5vp+5ppr9jvv63tdYzEMw0BEREREXKwF3QARERGRG406SCIiIiJu1EESERERcaMOkoiIiIgbdZBERERE3KiDJCIiIuJGHSSRG9CYMWPo1KkTnTp1onr16rRp08Y1feHChT/9fIZh0Lt3b1JTU6+4zIEDB6hSpQoffPDBX2m6iMjfgkXXQRK5sbVs2ZIpU6ZQo0aNa36O3NxcqlWrxs6dOwkNDb3sMiNHjiQrK4vdu3ezfv16/Pz8rnl9IiK+zr+gGyAif95PP/3E66+/TmpqKna7nd69e9O5c2fS09N58cUXOXHiBFarlRo1ajB69GhefPFFAHr06MGsWbOIiooyPV9aWhqrVq1i6dKl9OvXj3Xr1tG2bVsAbDYbY8eOZfPmzfj5+VGvXj1GjhwJcNn57777LpmZmYwYMQKASZMmuaYffvhhSpQowZEjR+jRowcxMTFMnDiR7OxsEhMTadq0Ka+99hoAGzZsYMqUKRiGQZEiRXj11VdZt24dJ06cYOzYsQDs2LGDt99+m8WLF3vldReRfw51kER8jM1mY/DgwUycOJGYmBhSU1P517/+RcWKFTl8+DA5OTksX76c3NxcRo0axcmTJ3nzzTdZsWIFCxYsuGyCtGzZMipVqkSFChXo3Lkzc+fOdXWQ5s+fz6FDh1ixYgUBAQE888wzrF27lqSkpMvOv5qbbrqJ1atXAzB48GCGDBlCvXr1SE9Pp2XLlvTo0YObbrqJYcOGMX/+fGJiYlizZg0TJ07k1VdfpW3btqSmphIaGsonn3xCt27dru8LLCKCOkgiPufIkSP8+uuvDBs2zDUvJyeHH3/8kYYNGzJlyhR69uxJ48aNefzxxylbtiy5ubm/+5yLFi3ikUceAaBTp05MnjyZvXv3UrNmTbZu3cr9999PUFAQAP/3f/8HQN++fS87f9KkSb+7rrp167r+P27cODZt2sS0adM4evQo2dnZZGRkcPToUapWrUpMTAwAbdu2dXXYmjRpwsqVK2nbti3btm1jzJgxf/i1ExH5o9RBEvExDoeDYsWKsXz5cte8xMREQkNDCQoKYv369Wzfvp1t27bRq1cvXn75ZVq0aHHF59u+fTvHjh1jxowZzJo1C4DAwEDmzp3LhAkT8PPzw2KxuJZPSkrC4XBccb7FYuHSoY02m820viJFigDOgePdunWjevXqNGnShPbt27Nnzx4Mw8Df3/zR5HA4OHz4MDExMfTo0YM333yT3Nxc2rZtS+HCha/hVRQR+X06i03Ex1SsWBGr1crnn38OwKlTp+jQoQMHDx5k3rx5jBw5kiZNmvDCCy/QsGFDfvrpJ1dn5nJJ0sKFC+ncuTObNm1i48aNbNy4kalTp7J27VoSEhJo3LgxK1euJCcnB4fDwciRI1m7du0V54eHh7N//34MwyAjI4OtW7deto5z585x8OBBhg4dSuvWrTl16hQnT57E4XBQq1YtDh8+zJEjRwBYt26daxxV/fr1sdlszJ07V4fXRCTfKEES8TGBgYFMmzaNN954g+nTp5Obm8tzzz3H7bffzq233srOnTtp3749hQoVonTp0jzyyCNYLBbuvvtuunbtyvTp07n11lsBZ/K0YcMGUxoFcNddd1G9enXmz5/P4MGDiYuL44EHHsAwDBo1akSPHj0wDOOy89PT0/n666+55557iI6Opnbt2petIzw8nMcff5xOnTpRrFgxihcvTu3atfnll19o0KABY8eOZejQodjtdkJCQhg/frzrsQ888AAbNmygYsWK+fdCi8g/mk7zFxGfYrPZ6N+/P126dKFNmzYF3RwR+ZvSITYR8RkHDx6kcePGhIeH07p164Jujoj8jSlBEhEREXGjBElERETEjTpIIiIiIm7UQRIRERFx45On+d8b3regm1BgjFvLFHQTCpTl2KmCbkKBsgQHF3QTCk7Rf3DtADm2qy/zN+YI+2dv///tGu3V9TniK3ttXdbow15b15+hBElERETEjU8mSCIiIpJ/HDi8tq4bNam5UdslIiIiUmCUIImIiIiJ3fBegnSjdkSUIImIiIi4UQdJRERExM2NmmyJiIhIAXGgXyFTgiQiIiLiRgmSiIiImHjzNP8blRIkERERETdKkERERMTEbmgMkhIkERERETdKkERERMREZ7EpQRIRERHxoARJRERETOxKkJQgiYiIiLhTgiQiIiImGoOkBElERETEgxIkERERMdF1kJQgiYiIiHhQgiQiIiIm+iU2JUgiIiIiHtRBEhEREXGjQ2wiIiJiogtFKkESERER8aAESUREREzsCpCUIImIiIi4U4IkIiIiJjrNXwmSiIiIiAclSCIiImJix1LQTShwSpBERETEJ6xcuZJ27drRunVrFixY4HH/pk2buO+++7jvvvt47rnnyMjIACA1NZV///vftG3blh49epCYmHjVdamDJCIiIiYOw3u3PyohIYFJkyaxcOFCli9fzieffMLPP//suj81NZXhw4czadIkVq5cSUxMDJMmTQJg8uTJ1KtXjzVr1tClSxdef/31q65PHSQREREpMKmpqZw8edLjlpqaalpu69atNGzYkGLFihEcHEybNm1Yu3at6/7jx49TqlQpKlasCECLFi344osvAPjqq6+47777AOjQoQObN2/GZrP9brs0BumiBq1r0GfUAwQE+nPswEkmDZpLZtqFyy773Lt9OH7wFIvfXQdAYKEAnh7Xncq1K2C1Wjm46yhThy4k58Lvv/g3kgZ3VqLP060ICPTj2E8JTBqzgsyMHNMyLdvWoMsjjTGA7As23hu/hp9+jHPdX6RoEONn9mHia8tN831Bg9Y16PNSZwKC/Dm2/xSTBs8lM/0K2/+d3s7tP3U9cHH7v/3b9rdwcNcxpg7zre1fv+Vt9BnWwfn+P3iayUMXkZmefdlln53QnV8OxbF45peueR/vGUNSfIprevGMjXy5bFe+t/t6qd+0Cn2G3OOs/3A8k19aQmbGFep//UF++TmBxR9+DYDVauGplzpSo/7NAOzcfIhZ49Z4re1/Vf3mMfR5vp1z3z8Ux+QXP7vytn+7K78cjmfxB5tc89r3aMS9/2pAYFAAP+8/xeQXP8WWY/dW86+rBndVos+AuwkI8OfYzwlMenW5x/ugZduadOl5J4ZhOD8Hx63hpx9PF1CL8483xyDNnTuXd99912P+gAEDGDhwoGv6zJkzREREuKYjIyPZu3eva7pChQrEx8dz8OBBYmJiWLNmDUlJSR6P9ff3p2jRoiQnJxMVFXXFdilBAsKKF+XZd3vzWq9pPHHHSOKOJ9Fn1AMey5WtHM1by56jyf11TfMffrY9fn5+PNXkVfrf9QpBhQLpOqStt5r/l4UVC+a5UZ14bdinPPHQVOJPneexAXeblilTvjhPDGrNiEELeKrHDBZ+sJlRY7u67q/fuCL/N6cvZSuU8Hbz/7Kw4kV59v968Vqf6TzRcBRxvyRefvtXiuatpc/SpFM90/yHh7TDz9/KU81epX/T0QQVDqDrMz60/cOL8Oz4hxnz5Gz6tniD+BNn6TP8Po/lylaM4s1FT9GkQy3T/NK3RJKWksmAtuNcN1/qHIXdVIRnX3+QMc8spG/7ScT/mkyfZ9t4LFf2lgjenP04Te6tYZrfsmNtSlcoQf9OU3iq8/9Ro97N3NWmurea/5eEhRfh2be7Mubpj+h7zzjiTyTTZ2g7j+XK3hrJm/OepEm7mqb5je+pTsdH7+TFnjPp13YCgUEB3N+nqbeaf12FFQvmuZfv57Whn/DEg+8Qf/Icjw28zOfgM/cwYuA8nuo+3fk5OL7rFZ5R/qhevXqxYcMGj1uvXr1MyxmG5/E4iyWvIxcaGsrbb7/NyJEjefDBB4mMjCQgIOCK67Vaf78L5LUO0pEjR3jvvfcYNWoUr7zyCu+99x779u3z1up/V50W1Ti85zinj54B4PPZX9Gyyx0ey933eAvWL/yGLW4f/vtiD7NowucYhoHDYfDz3hNElinulbZfD3Ua3sqhA6c4/WsyAKsW76Sl2x8BW04uk8esJPlsOgCHfzzNTcWL4u/vfAvd3+0Oxo9extnENO82/jqo0+I2Dn/3S972/3ATLR+60vbfypbl35rm74v9iUUTL9n++34lsky4V9p+PdRpGsPh709w+rjzm9aqed/Qwu1LAECHnnex/rMdbFn1nWn+bXUr4LAbvPXx07z3vxfoPrgNVqvvnAFT586KHP7hJKd/OQvAqo+308KtEwjQ4eGGrF+6iy1rzZ9bVquFQsGBBAT6ExDoj3+AH7bsXK+0/a+qc1dlDu/9ldO/XNz2C2Np0bG2x3IdHmnM+sU72bJ6r2l+q851WTJ7M+kpWRiGwbujFrPRhzrHl6rT6FYOHTid9zn43520bGvuENpy7Ex+bTnJSRc/Bw/89jno5/X25jc7Fq/dQkNDKVOmjMctNDTU1KaoqChXIgTOVCgyMjKvzXY70dHRfPbZZyxevJjq1atTtmxZwJk2/fbY3Nxc0tPTKVas2O++Bl7pIC1YsIBnn30WgBo1alCtWjUARo4cyezZs73RhN8VUfomEk+dc00nnj5HkdBggkMKmZZ7b9giNny6zePxu788wKkjCQBElgmnc7+72bLcdz4kIqJCSUrIO9abeCaVIkULEVwk0DUvIS6FHd/85Jp+ckgbtm0+RG6u83JiIwYt4Md9J73X6OsoolQ4iaeSXdPO7V+Y4KJu23/4IjZ8dpnt/9UBTh1xdq4iy4TT+clWbFnhO9u/RKliJMadd00nxZ2/WH+Qablpoxazccm37g/Hz9/Knq8PMbLndIZ2eYc6zarQ0YdShBLRYSRecngwKSGVIiGFCC7iVv/rK9m48jv3h/PFst2kp2Qx78vhLNj0IqdPnGX7Vwfzvd3XQ4mSbts+PoUiIZfZ9qOXsXHZbo/Hl7k5gmLFi/La7Cd4b9Wz9Bh0D+mpWfne7vwQERVmOkyc9zmY91okxJ1nx9eXfA4+24Ztmw6Rm+ubhxR9TePGjYmNjSU5OZmsrCzWrVtH06Z5nzUWi4XHHnuMhIQEDMNg9uzZtGvnTESbNWvGsmXLAFi9ejX16tX73XQJvDQG6aOPPmLZsmUULlzYNL9Pnz507tyZxx57zBvNuCLLFWI2u/3PXUu04u3lGDXvKVbM+pId6/Ze/QE3CKvl8t/27Zf5MZ6gQgE8/3InIqLCGDFofn43zSssV0g77I5r2P5zf9v+N0Y6+kf8me1/OWsX5XUabTlZLH3/Kzr1acqyS8ap3MiulHb90e3f46lWpJzLoHvTNwgM8mfUO4/yQO+7WDLn6+vZzHxxxdr/4Gefn7+V2ndW4tV+c8jJzuW5sV3p/WxbZry+4no20yuuvB94vhZBhQJ4fnRnIqJCGTHg7/E56M5h3HgpcFRUFEOGDKFnz57YbDYeeughatasSd++fRk0aBA1atTg1Vdf5YknniAnJ4dGjRrx+OOPAzB48GCGDx9O+/btCQkJYfz48Vddn1c6SP7+/uTmekbOFy5cuGoPzhsST54lpu7NrukSJYuRdi6D7Myc33mUWbMH6jNgXA+mvrCQrxbvyI9m5pszCSnEVC/tmi4REUpaShbZboOMI6JCeXXiw5w4nsQL/eeS4yOHEa4m8VTyX9/+neszYGx3pg5f5Hvb//Q5qtQu75ouER1G2vkMsrP+WP0tH6jH0QOnOH7QOTDfYrH41DfqM3EpVKlZ1jVdIiqUtJRMsrP+2CD7xq2rMe31leTa7OTa7HyxfDd33VPdJzpIZ06fp8rt5VzTJaJCSTv/x2tPPpPK1nU/uAZ1b1y+m+4DWudLW/PbmfgUYqqXcU2XiAhxvg/cPwejw3h1UndOHEvkhSfn/G0+B33Fb9c4utT777/v+n/z5s1p3ry5x+OKFSvG9OnT/9S6vHKIrV+/ftx///289NJLTJkyhSlTpvDSSy/RpUsX+vXr540m/K5dXx4gpt4tlLrFeSyzfZ9mxK7xjNKv5K6Odej/Zjf+8+Akn/vjCLBr2xFiqpehVFnnuJn2D9YjdrP5EEFIaCHGz+jN118e5M0Ri/9WHwq7vjxATN1Ltn/vP7n976tD/ze68p8uk31y++/efIiY2hUodXGAfbtH7iR23Q9/+PEVqpTk0efaYbVaCAwK4L5eTdi8ck9+Nfe62/3NT8TULEep8s5xg+26NiB2449/+PE/HzhF04tj9vz8rTRsUZWD3/+aL2293nZvOURMrXKUKn9x23dvROwX+//w479es48mbWsSGOT8rt2odXUO7/ON2t3t2naEmBqXfA4+VJ/YTYdMy4SEFmb8zD58vfFH3vzPf/9Wn4PiySsJ0n333UeDBg2IjY3lzJkzGIZBvXr1GDhw4O+eYuctKUlpTBzwIS/N6Yd/oD9xxxIZ1/8DKtUqzzNTevF0s1d/9/F9Rj4AFgvPTMkbcX9g+89MfWFhfjf9ukg5l8mEV5cz8q0u+Af4EXfyHONeWUqlqiUZ8lJHnuoxgw4P1iciOow7W8RwZ4sY12OHPfURaSm+OebgNylJaUwcNIeXZj/p3P7HExn31Gzn9p/Uk6dbvPa7j+8zsrNz+0/q6Zp3YMfPTB22KL+bfl2knE1n0vMLGTG9D/4B/sSdSGL8MwuoVLMsg9/uxoC243738QsmreWp1x5i2vph+Pn7seXz70yH3W50KckZTHrpv4yY1N35/v81mfEvfkalaqUZ/FpnBjzgefrxpWa+9Tn9R3Rk5qohOBwOvtt2hM985PBiSnIGk4Z9yoh3H3XWfuIs44d+TKXqZRj8RhcGdJz0u49ftWArRYsF887yZ7BaLfy8/xSz3lzppdZfXynnMpgwehkjx3a9+DmYzLhRS6lUtRRDRnbkqe7T6fDQFT4H+8/1+c9Bd/qpEbAYlztv7gZ3b3jfgm5CgTFuLXP1hf7GLMdOFXQTCpQlOLigm1Bwiv6DawfI8Z3rauUHR9g/e/v/b9dor65v94lyV1/oOqlT7oTX1vVn6EKRIiIiYmLXZRL1CoiIiIi4U4IkIiIiJjfiaf7epgRJRERExI0SJBERETHRWWxKkEREREQ8KEESERERE7uh/ESvgIiIiIgbJUgiIiJi4lB+oldARERExJ0SJBERETHRWWxKkEREREQ8KEESERERE53FpgRJRERExIM6SCIiIiJudIhNRERETBwapK0ESURERMSdEiQRERExsSs/0SsgIiIi4k4JkoiIiJjoNH8lSCIiIiIelCCJiIiIiX6sVgmSiIiIiAclSCIiImJiN3QdJCVIIiIiIm6UIImIiIiJroOkBElERETEgxIkERERMXHoOkhKkERERETcKUESERERE41BUoIkIiIi4kEdJBERERE3OsQmIiIiJrpQpBIkEREREQ8+mSAdfC2moJtQYKw3ZRd0EwqUYdxc0E0oUP6B9oJuQoHJjS9c0E0oUI7CjoJuQoF65s71Bd2EfxT9WK0SJBEREREPPpkgiYiISP6x60KRSpBERERE3ClBEhERERMHOotNCZKIiIiIGyVIIiIiYqIxSEqQRERERDwoQRIRERET/VitEiQRERERD0qQRERExMSh32JTgiQiIiLiTgmSiIiImGgMkhIkEREREQ/qIImIiIi40SE2ERERMXHoQpFKkERERETcKUESERERE7t+rFYJkoiIiIg7JUgiIiJiojFISpBEREREPChBEhERERONQVKCJCIiIuJBCZKIiIiYaAySEiQRERERD0qQRERExMSuBEkJkoiIiIg7JUgiIiJi4tBZbEqQRERERNwpQRIRERETjUFSgiQiIiLiQQmSiIiImDgMjUFSgiQiIiLiRh0kERERETc6xCYiIiImduUn6iD9pkX5m3mhYRMC/fw4eDaRYRvXkW7LMS0z4s5mtLu1MinZFwA4eu4cA9atIsBqZXTTltQvWQaAr04c482tm3EYhtfruFYtSt3K0NtbOOs/f4bh2z4nPddc/39qt6JduRjO51ysP/Usg75ZRqDVj1F176FxdHkybDlsOPUzU/Ztxneqd9b/Qq3mzvrPnWHYttUe9Y+o05K25WLytn/aWQZ+vZxAqx8v12tNo6jyZOY665+8d4tP1d88uiLP1WhBoJ8/h84n8J9vV3nUXzk0glG17yUkIAi74WDkrtXsPx/vuj8kIIiFzXvy4rer+OFcnLdLuGYtKtzMC40u7vtJiQzbcJl9/65mtK1YmZQLF7f9+XMMXLsKCzCscVNaVLgZh2FwPOU8IzauJ/lCVgFUcm1alr2FF+o3JdDqx8HkRF7Ystaj/pfuaE67m6tw/rf3fkoyAzauBKBthco8XesOAq1+nEpPZcim1a7lfMEv32ayfd557DaD4hUCaT6gOIHB5s7BsW2Z7Fx0HosFgopaafZ0ccJKBrjuT0/MZcmweLpMLknhUD9vlyD5RB0kILxQYca1vJeHlizieMp5hjdqwrBGTRi5eYNpubrRpRi47nN2x582ze9ZozbhhYK5Z9EcrBYLnz3QjQ4Vq7Dip4PeLOOahQcF83bDDvxr/UccTzvHsFoteKFWC0Z9+z/TcnUiSjPom2XsTjplmv9UtcaULhLKvZ+/j81h5/UGbXmkcl3mHd7lzTKuWXhQYcY2ak+XdfMu1t+cF2q3YNROt/pLlGHQ18uvUH8Y934+C5vDzht3tOXRynX5yFfqDwzmrfr30fXLOfySfo6hNVryfI2WvLJnrWuZQn7+fNi0O//5dhWb4o/QqlRlJtxxP/f+bzoAzaJvZUSteyhdpFhBlXFNwgsVZmyre+nyX+e+P6xxE15o3IRRm8z7fp3oUgxa67nv/+u2GlSPjOS+j+eT47AzvHFTRjRpxnPr1+ILwgsVZlzTe3lw5UKOp55neP2mDK/flJe2fmFarm5kaQZuXMmuM+b6a5SI4tXGrei8YgEn01MZeUcLhtZrwohv1nuzjGuWlWLny3fOcv+b0RQrFcC2uefY9tE5mvYr7lomN9vBhklJdJlckrCSAXy/IpVvZp2j3chIAA59mc7OhefJTLYXVBn5QoO0NQYJgCblyrP3TDzHU84DMP+H7+lUuappmUCrH9VKRPLvWvVY0/VRpt17H6WKhgDwwfe7GLBuFQZwU6HChAYGcd6HvkE2KXkz+87GcTztHADzf9pNpwrVTMsEWv2odlM0fas25PO2j/PeXQ9QKjgUgOrh0az65QA5DjsGsP7kYdqWjfF2GdesSclb2Guqfw+dKtxmWibQ6ke18Cj6Vr2D1e0e470mnfPqLx7NykvqX/frYdqWq+LtMq7ZXdG3sO/caX5Jd9a/8MguOpavbl4m6hZOZJxnU/wRADacPszgbUtc9/es1IAXdqzgTFaa9xp+HXjs+/u+p1OVy+z7EZH0rVOP1Q8/yntt8/b9w8lJvPnNZnIczj+O+87EUzok1LtF/AVNS1dgb2I8x1Mv1v/jd3Sq6Pnev614JP+uWZ81nXsxvVUnShVx1t+54m18cmgfJ9NTAZi8+xum793h3SL+gl+/yyKyYhDFSjnToNvuDeHnzRkYl6T/hgMwICfDAUBulgO/AGfnISM5l+PbM2k3KtLrbZf8pw4SUKpoKHHpeR/scelphAYFUTQg0DUvskgRtp46wdhtW2j7yTz2JMTxfrv7XffnOhwMa9iETY8+TlJWJjvizCnDjaxkcChxmamu6fjMVEICC1HU/5L6CxclNuE4Y7/7kvZrPmDP2VPMaPoQAN+dPU378rcR7B9AgNVKxwrViCxc1Ot1XKuSwSEe9Ydepv6t8b8w7ruvaLd6Nt8lnWZmswcB+D7pNB3KV/XZ+qMLu23/rFRCAsz13xxSnKQL6bxRrwNLWj3GnKY98LPkfXw8vmUR3yX7znv+NyVDQolLy9v34y+37xctwtaTJxi3dQvtFs3ju/g4ZnZw7vt74uPYn3gGgNCgIAY2aMTqnw97t4i/oGSREE5nXPLZl5FGaKC5/qjgosTGneDtnZtpu3Que86cZlbrzgDcEhaOn9XK+63vZ03nXrx2Z2sy3A7P3cgykuwULZF3SKxoCT9yMg1sWXkdpIDCVpr0D2fp8Hg+6nOSH1an0bCXMyktEu5Pm+GRhJcN9HhuX+fA6rXbjcorLTt9+vTv3gqa5QpJot1wuP5/Mi2VPquWcvS881v2zD3fUi4sjDKXfFt8e9sWbp81lZOpKbze7O58bfP1ZL3Cb+7YL/kWdTIjhce++pRjackAvP/jdsqF3ESZImHMOBDLTymJLL6nF/Nadmd34ilsDt+Jm61XeANcrv6jF+ufeUn90w/E8tP5JBa36cn8Vg+zO+mkK1HwBX+kfn+LlWbRFfnk6G4e2DCbeT/vZFaTbgRafXu8hfWP7PupqTy28jL7fmjevl8uNIxPHujKt6dP8dHe7/K1zdfTH9n2v6an0Pt/izma4qx/xr6dlAstRtmiYfhbrdxd7lb+8/V62i2dS2JmBm/d1cYrbb8eLtnMJpf0/Tl7PIddn6TQ9Z1S9PywDHW6hLHu7URTyiR/T14Zg/Tkk09y/PhxIiMjPd5UFouFDRs2XOGR3nE6LY3aUSVd09FFi3L+QhZZubmueTHFS1C1eARLD//ommfBQq7DQd3oUiRnZXEs5Ry5Dgf/PbifV5q29GoNf8WpzBRuL1HKNR1VOITz2Vlk2W2ueTHFIogpFsWy4z+45llwJmdhgYWZ9eN23tyzEYD25aq6Dlf5gtMZqdQqnld/dPDl6696UxRLj12+/vd/3M4bv9Vfviq/+FL9mSncHn7p9g/lfI65/jMX0jmalsT3yc4vNBtOH+aNeu0pW6QYR9LOer3N18vptDRq/ZF9v0SwrjzJAAAgAElEQVQESw+57ft251/XhqXL8s69HZi5eyfv7/nWe42/Dk5npFEr8pL6i4RcrP+S9354BFXDI1j68wHXPAsWbA47CRnpHEpOIjErA4BPD+9jUfuu3ivgLyoa4ceZn7Jd0xln7QQVtRJQKK+H9Ot3WURXDXINyq7WNoSts89xIc3xtx6QbdcYJO8kSIsWLeLmm29m7NixbNy40XQr6M4RwJZfj1MrqiQVwpyxaY9qt7P+2BHTMg7D4JUmLV2J0SPVb+fg2UTiM9JpXKYcI+9qjp/FggXoVLkqsSd/9XYZ1+zruGPULl6aCiE3AdCjUh2+OGk+TOAw4OV6rSlTJAyARyrV4eD5M8RnpXF36Uq83qAtAMH+ATwe04AVx/d7t4i/YEvcMWqXyKu/e6XarD/5k2kZh2Hwct273epPJD4rjdZlKvHGHXn1PxHTgOXHfKf+rxOOUqt4acoXddb/8C112HDKvP03xf1M6SLFqFYsGoD6JcphGPBrxnmvt/d62nLiOLWj8/b97tVvZ/1Rz33/5aYtXYnRIzVu52CSc9+vE12KGe078tz6NT7XOQLYfPI4tSNLUSH04mdfzO2sO/GzaRmHYTC6USvKFnW+9x+tWouD5xKJz0xn9fHDtCh7C8WCCgHOM9r2JsbjK8rWKkzCoWzOn3Z2CA/8L40KDQqblom4JZDTP1wg87wzFT6+PZOQSP+/dedInCyGl3LCvXv38tlnn/Haa6/95eeqMHXCdWiRWfPyNzOs4V0EWP34JfU8z36xlnKhYbzd8h7afTIPgPsrV6V/nQb4WSzEZaQzbOP/OJ2eRoDVyqi7WnBH6TI4DINv404x5ptNXLjkW+j1Yr0p++oLXYPmpW5l6O3NCbD6cSL9HM/FrqRc0WK8eUd7Oqz5AIBOFarR77bG+FksxGemMXz755zOTMXPYuH1Bm2pVbw0fhYrHx/ZwwcH82egppFP32qal7qVF2o1c27/9PM8t9VZ/1t3tKP9mtkA3F+hGv2qNXLVP2zbalP9tUuUctb/83fMyqf6/QPz59Bds+hbea5GSwIvbv+hO5ZTtuhNvFGvPR3XzwKcnaJhNVtR2D+AHIedMXvWseus+YvAl+0GMDB2cb6c5p8bX/jqC12D5uVv5oXGF/f9lPM8t34t5cLCeKvlPbT/+OK+X6Uq/eo69/349HSGbXDu+/M6PUTNqGhOpqa4nu/X1BT6rV5x3dvpKHyF40F/UYsyNztP8/dzfvYN2bSaciFhvN3kXtotnQs4B2P3r9kAP6uVuIw0Xti81jV26ZGqtehZtRZWi4VT6akM3bKWM5kZ172dz9yZP2fG/fJtFjvmn8OeaxAaHUDLwcVJTchl07tn6TLZmaz+sDqNH1an4udvIaiolbv+HU54OfO4o+n3/0Kvj8rkW8dpSNV1+fK8VzJ4z8NeW9eU2ou8tq4/w2sdpOspPzpIviK/Oki+Ir86SL4ivzpIviC/Oki+Ir86SL4ivzpIvkIdJO/TdZBERETExGHcuGeXeYteARERERE3SpBERETExH6Fy7/8k6iDJCIiIj5h5cqVTJs2DZvNRu/evenRo4frvh9//JHhw4e7ppOTkwkLC2PVqlUsW7aM8ePHU7y482dkmjdvzpAhQ353XeogiYiIiMmN+FtsCQkJTJo0iSVLlhAYGEi3bt244447qFixIgBVq1Zl+fLlAGRlZdGlSxdeeeUVAPbt28fw4cPp0KHDH16fxiCJiIhIgUlNTeXkyZMet9TUVNNyW7dupWHDhhQrVozg4GDatGnD2rWX/2HoGTNmUL9+ferVqwc4O0jLli2jY8eOPP/886SkpFz2cZdSB0lEREQKzNy5c2nVqpXHbe7cuablzpw5Q0REhGs6MjKShIQEj+dLTU3l008/ZcCAAa55ERERDBw4kOXLl1OyZEleffXVq7ZLh9hERETExJun+ffq1YvOnTt7zA+95PcOgcv+/p3lMr8nuHLlSu6++27XeCOAqVOnuv7/xBNPcPfdV/+9VCVIIiIiUmBCQ0MpU6aMx829gxQVFUVSUpJr+syZM0RGRno83xdffEG7du1c02lpacyZM8c1bRgG/v5Xz4fUQRIRERETBxav3f6oxo0bExsbS3JyMllZWaxbt46mTZualjEMg/3791O7dm3XvODgYGbNmsX3338PwPz582nduvVV16dDbCIiInLDi4qKYsiQIfTs2RObzcZDDz1EzZo16du3L4MGDaJGjRokJycTEBBAUFCQ63F+fn5MnjyZV155hQsXLlChQgXGjh171fXpt9h8jH6L7cY79dSb9Fts/1z6LTb9Fps39dnZx2vr+rD+h15b15+hQ2wiIiIibnSITUREREz0Y7VKkEREREQ8KEESERERkxvxp0a8TQmSiIiIiBslSCIiImLyZ65P9HelBElERETEjRIkERERMdEYJCVIIiIiIh6UIImIiIiJroOkBElERETEgzpIIiIiIm50iE1ERERMNEhbCZKIiIiIByVIIiIiYqILRSpBEhEREfGgBElERERMNAZJCZKIiIiIByVIIiIiYqIESQmSiIiIiAclSCIiImKiBEkJkoiIiIgHn0yQprSdV9BNkAISbk0v6CYUKBt+Bd2EAmP/h/94ZqvC9oJuQoE668go6Cb8oyhBUoIkIiIi4sEnEyQRERHJP7qSthIkEREREQ9KkERERMREY5CUIImIiIh4UAdJRERExI0OsYmIiIiJDrEpQRIRERHxoARJRERETJQgKUESERER8aAESUREREyUIClBEhEREfGgBElERERMDCVISpBERERE3ClBEhERERP9WK0SJBEREREPSpBERETERGexKUESERER8aAESUREREx0FpsSJBEREREPSpBERETERGOQlCCJiIiIeFAHSURERMSNDrGJiIiIiQZpK0ESERER8aAESUREREw0SFsJkoiIiIgHJUgiIiJiYhgF3YKCpwRJRERExI0SJBERETFxoDFISpBERERE3ChBEhERERNdB0kJkoiIiIgHJUgX/bjDzpoPbeTaoOTNFro8E0ihIuYe9A/f2Fk334bFCoWLWugyOIDipZx9zNFdswgtkbd8swf9qdPSd17ef3r9e7fD4g+t5NqgzM3Qe4iDwkXMy+z+BpbPs2K1QHBR6DXEQWQpcNhhwVQLh/c5669R36BLXwOLD30B27cdln8Itov1PzIEj/q/+wZWzQPLxfofGQIRpWDma5B4Om+5pHioVBOeGu3dGq7VD9sNVnxokGuD0jdD9yEWCru997//xuDzeQYWq7P27s9YiCjlXGbzSoOtaw1sOVCuovPxAYG+s/G/ioVJMyHHBlVugTHDoKjbtp+/GBYshUJBcEs5GDkEioU671u4FP77OWRnQ7UqMOYFCAz0fh3XamusHzNmBZJjs3DrLQ5eHHqBIm71/3dJAIuXBRAUaFC+vIPnBmcTGgrp6fDWuEL8csKCYVi4t42NRx62FUwh15mug6QECYD08wafTszh0ZcCeWFWIYpHW1nzoflNbss2WDQuh54jAxkytRC3NfRj+XTnMmdOOigcYmHI1EKumy91Dv7p9aedhw8nWHlqpIPXP3AQEW2weLb5wyEnG2a97Vzm5WkObm9ksGiac/eJ3WAh4aSF0dOd9x3eZ2HXloKo5NqknYePJsC/R8LoD6BENCybbV4mJxs+fNu5zIhpULMRfDrNed9v80ZMgx7PODsQ3Z72fh3XIu28wfyJBk+MtDDqAyvFS8KKD83nN+dkG8wda9B3lIUX37NSo6GF/05zLvPd1wabVhgMfMvCiBkWcnLgy6UFUcm1ST4PI96CKa/BmvlQphRMmGFeZvtumLUIPpwISz+Apg3h5fHO+9ZthgVLYPZEWDkXLmTD3M+8X8e1Once3hgbxJjRF1j0USalSjqYNjPItMzuPX4sWBTAlAlZzJmVRaM77IydUAiAWbMDiYhwMO/DLN6flsmy5QH8sF9/Vv8utCWBw7vtlK1sJaK08+Vo2MGPPV/aMS65EITD4fz3Qobz35wsA/+L35J+OeDAaoXpw7KZ2P8C6xfYcNh95yIS//T69++2UKEKRJV2TjfvYLB9o8V0HZDf6s/KdP6bnQUBAXn3ZV9wpi+5F2/+PvQN+sfdUKEKRF6sv2kH2LERj/oNzPX7B5ifJ9cGc8dDl34QHumVpv9lB3dD+coQWdrZIW7S3sLOjZje+8ZvtV9872dn5W3fHRsMWj1goUiIBavVQreBFhq08nIRf8E3O6F6DFQo45x+uBOs+sK87fcfhkZ1IfriNm3dFL7c6kycVvwPend1pklWK7zyHHS8x/t1XKudO/2pWsVB2TLOgjt3srF+g7+p/oOHrdSraycywjmzWZNcvon1w2aDwQNzeLp/DgBnky3YbFCkiO989v0ew/De7Ublta/5X3zxBXFxcTRr1oxy5cq55n/yySd07drVW824rJQkg7CIvMQgrISFC5mQnQmFLkatQYUtPDAggKnPZhMc6vzQfGqC85uGww6Valtp/3gAthyY/XIOhYLtNOnsGynKP73+5EQIL5G3l94UAVmZztfgt8NMhQrDIwMN3hpipUiIs8MwfKKz13Rna4NvN1sY2sOK3Q7V6kCthgVRybU5lwg3lcibLhYBFzLxqL/7QBg/BFf9z080P883ayEsHGrd6b22/1XnEp31/uZytQcVttBtIEx81iA4xMBwwLMTnfvLmVOQlgJTRzhIOQu3Vof7n/CdQxPxZ6DkJZ3ZqAhIz7CQkWm4DrPVqOo8xHYqHkpHw9I1YLNZOJ9qcPxXqBEDfYfCmSSoWxOe71cwtVyLhEQLkZF5+35EhEFGhoXMTFyH2W6LsfPfJQHEx1uIjjZYvTYAm81CSqqFEsUN/P3g1deD+GqTP02a5FKu7A38F1/+FK8kSOPHj2f+/PkcP36cbt26sXz5ctd9H3/8sTea8LsMx+XnW/3y/h93zMEXC3N5fkYQIxcUpmW3AOaNycEwDO5o60+n/oH4B1ooXNRC087+/LDV7p3GXweq//LzL63/5DFYucDCqzMdTFjkoP3DBtNes2IYsGK+hZBiBhM/djBugYOMNPjff33nj6TjD9R/6hisXgCjZsJbi+Deh51jjy799rdxKbTrnr9tvd6u9O3VXLvBmgUGI2ZYeGOhlTbdLMx6zcAwDOy5cHC3wWP/sfDCOxYy02DlHN/5A3nFbX/JX4b6t8NTvWHgS/DQv51j0MJCDQL8wZYLW7+FSa/AZzMhJRUmz/JGy6+PK+77l9Rf63YHj/XM4T+jCvH4k4WxWAxCQw0C/PO286gR2axankFaqoU5H/lQfPw7DMPitduNyisdpE2bNjFr1ixGjhzJwoULmTJlCmvWrAHMUXZBKRZpIS05rx2pSQaFi0JgobwNd3iXnQq3WV2Dkht38CP+F4PMVNi1IZe4Y3l7mmGAn2+EJ4DqD4+ElOS8Ws8nQXBRg6BCecvs/9ZCxWoGkaWc0y3vMzj1C6Snwu5vLNx1j4F/AAQXgcatDQ59f+Pu9O6c9edNO+vHVP+Bb+GWas5B2QDN74PTv0BGqnP615/BbncOzvYlN0VA6iW1p7hqz9t+P+76rXbnvKaX1B5WHG5v7BzU7R9goX5LC8d+9HYV165kFCSezZtOSIKwEIPgwnnzMjKdnaQls+C/M+GeZs75xUIhsgTc3cQ5qDswAO67B77f790a/oqoKIOzZ/O2dVKihZAQg8KX1J+ZCbVq2Zk9M4sPZmTRvKnzy19oKGzf4UdSkvPxwYXh7la5HPpJI1f+LryyJQ3DwHLxlJ4KFSowY8YMXn/9dbZv3+6aX5Aq1/HjxEEHiaecf+S3rbZTrZGfaZnSFa0c3ecg7ZyzI7E/1kF4lIUiYRYSjhusm+ccd2PLNti6Mpfbm/p5rOdG9U+vv1pdgyMHIeGUc/qrzy3UamTuuJerZHB4n4WUc87pPVuhRBSEhEH5igY7Nzvfx7m58N02C7dULfiO/x9VtS4cO+g8XASw5XO4vZF5mbKV4Kd9kHqx/u8u1l80zDl9eC9UqYVPnbkHztqPH4Qzp5zba8vnBjXca68IP++F1Ivv/e9joXgUFA2zUOsuC3u2GORkOxOlvbEG5St7u4prd2d9+P4AHD/pnP5kBbR0O0R6Jgl6PQPpF8dgTfsI2rdybus2zeB/XzkHZxsGbNjiHNPkKxrUs7P/Ryu/nnS+cZetDKDJnbmmZZKSLAx8pjAZF+ufMy+Au1vasFhg41f+zJ4biGFATo5zum5t30nP5fdZDC9EOO+++y5bt25l+PDh1Kzp/Iq5a9cuBgwYQE5ODrt27fpTz7f8aK3r3sYfd9hZO8eGPRfCS1ro9nwgZ+MM/jslhyFTnV+lt67MZevKXPz8oXCIhfufCiC6vJWcCwbL3rNx4qADux1qNvHj3l7+N0Tn74/ylfrDrenX/TkB9u6AJbOt5OZCZEl4bKiDpHiYO8nKy9OcHceNKyxsXGHB3985Dqf70w5KV3CmSAunWjjxswWrFWJqG/zr3wb++ZCi2cifjucPO5xnrtlzoURJ6D3Uebr+/EnOs9MAvloBm1Y408EiIdD1aShVwXnfoned44/y8xCb3cif73P7d1w8zf9i7T2HWkiKg4WTDV58z7nOTSsMNq90btPgEPjXUxZKVrDgsBusXQS7Nxs47M7OVLdBnpcJuB5aFc6fP7ybtjlP87fZoGxpeOs/cPI0jBznPGsNnGeqLVwKDgPq1ICRzzhP+bfbYfo8WLMR7A64rRKMft7zMgHXw1lHxvV/UiB2mx/T3w8kN9dC6VIOXnrxAqfjrLw1Log5s7IAWLw0gCXLAnAYULO6nWcHZxMUBGnpMH5iEEePWbFYoMlddh7vnWM6RHe9RJQ6df2f9HfUWPGy19a1r+ONeU0Qr3SQAGJjY4mMjOTWW291zYuLi2P27NmMGDHiTz1XfnSQxDfkVwfJV+RXB8kX5FcHyVfkVwfJV+RXB8lXqIPkfV4bKdKoUSOPeSVLlvzTnSMRERHJX7pQpK6DJCIiIuLBh841EhEREW+4AU4wL3BKkERERETcKEESERERkxv5Ao7eogRJRERExI0SJBERETFRgqQESURERMSDEiQREREx0UlsSpBEREREPChBEhERERONQVKCJCIiIuJBCZKIiIiYaRCSEiQRERERd+ogiYiIiLjRITYREREx0SBtJUgiIiIiHpQgiYiIiImhQdpKkERERMQ3rFy5knbt2tG6dWsWLFjgcf/Ro0d59NFH6dixI48//jgpKSkAnD59mh49enDvvffSv39/MjIyrroudZBERETExDAsXrv9UQkJCUyaNImFCxeyfPlyPvnkE37++edL2mzQv39/+vbty4oVK6hatSozZ84EYPTo0XTv3p21a9dSvXp13nvvvauuTx0kERERueFt3bqVhg0bUqxYMYKDg2nTpg1r16513b9//36Cg4Np2rQpAP369aNHjx7YbDZ27txJmzZtAHjggQdMj7sSjUESERERMy+exZaamkpqaqrH/NDQUEJDQ13TZ86cISIiwjUdGRnJ3r17XdMnTpygRIkSDBs2jAMHDlC5cmVGjhzJuXPnKFq0KP7+zi5PREQECQkJV22XEiQREREpMHPnzqVVq1Yet7lz55qWMy4zctxiyevI5ebmsmPHDh555BFWrlxJ2bJleeutt676uCtRgiQiIiIm3jyLrVevXnTu3Nlj/qXpEUBUVBTffvuta/rMmTNERka6piMiIihfvjw1atQAoEOHDgwaNIjw8HDS09Ox2+34+fmRmJhoetyVKEESERGRAhMaGkqZMmU8bu4dpMaNGxMbG0tycjJZWVmsW7fONd4IoHbt2iQnJ3Pw4EEANm7cSLVq1QgICKBevXqsXr0agGXLlpkedyVKkERERMTsBrwOUlRUFEOGDKFnz57YbDYeeughatasSd++fRk0aBA1atRg6tSpvPTSS2RlZREdHc3YsWMBePnllxk+fDjTpk2jZMmSTJw48arrsxiXOzh3g1t+tFZBN0EKSLg1vaCbUKBs+BV0EwqM3fhnB96tCtsLugkF6qzj6tet+TuLKHXKq+u7ZeEbXlvX0e7/8dq6/gwlSCIiImKi32LTGCQRERERD0qQRERExMznBt9cf0qQRERERNyogyQiIiLiRofYRERExESDtJUgiYiIiHjwyQSpWaGzBd2EAlPYElDQTShQufyzrwWT4sgs6CYUmDDrP/u9Pzu1XEE3oUB9/MS9Bd2EAvXFZi+vUIO0lSCJiIiIuPPJBElERETyk8YgKUESERERcaMESURERMw0BkkJkoiIiIg7JUgiIiJipgRJCZKIiIiIOyVIIiIiYqYraStBEhEREXGnBElERERMDI1BUoIkIiIi4k4JkoiIiJgpQVKCJCIiIuJOHSQRERERNzrEJiIiImY6zV8JkoiIiIg7JUgiIiJiYtEgbSVIIiIiIu6UIImIiIiZEiQlSCIiIiLulCCJiIiImc5iU4IkIiIi4k4JkoiIiJhpDJISJBERERF3SpBERETETAmSEiQRERERd0qQRERExEwJkhIkEREREXdKkERERMRM10FSgiQiIiLiTh0kERERETc6xCYiIiImFg3SVoIkIiIi4k4JkoiIiJgpQVIH6Tdfx1qZNiuQHBtUvMXBiKE5FC1iXubTJf58tsyfoECoUN7B0ME5hIVCejqMGRfILyesOAxo3yaXng/nFkwh12hzrIUp7/uRY4PKtxiMfsHuUf/CJVYWLbVSKBBuLm8w4hk7YaHO+5p18ieyRN6yvbvZad/ad/awLbFW3nnfH5sNKt1iMOoFm0f9Hy/x45OlfgRdrH/4MzZX/b95bmQAEcUNhj/jW9s/dpsfs2YFYcuxcMstdoYOvUARt/qXLAlg2bJAAoMMypdzMHjwBUJDITsbJk8pxKFDVhwOqFrVwTODLxAUVDC1/Fn/9G1/ZGc2mz9Kx26DiAr+3DsohKBg88GFw7HZfLMwHYvFQqGiFtoMDOGmkv7Ysg2+mJ5G/E82DAeUrBLA3f1CCAjynTOg7mhYkcefbE5AgD9Hj5xhwturyMzMMS3TqnV1/vVwQwwDsrNtTJ2yjsOH4vD3tzLgmTbUqFkWgB3bj/D+tI04HL7z2SdXpkNswLnzMGZsEG+Ozuazjy5QuqTBezMDTMt8u8fKR4v8eXdCNvNnXaDxHXbenBAIwIzZAURGGCz68AJzpl1gyXJ/9u33nZc2+TyMfNuPia/msnJeLmVKGUyeaW7/jj0WZi+08v6EXD77IJcmDR2MHu8HwLETEBoCn32Q67r5Uufo3Hl45e0Axr9qY+m8HEqXMnhnpvm7w849VuYs9Gf6hBw+/iCHOxvaGTPe/B6Zs8iPPXt9Z7v/5vx5C2PHFmL0K1l89FEGJUs5mPm+uXezZ48fiz4OZMKETGa9n8kdd+QyYWIhAObPD8Ruh1nvZ/LBrExysmHBwsCCKOVP+6dv+8wUB2unpHL/i2E8Mb04xaL92Dwnw7SMLdtg9YQU7n8xjN7/F07FBkFsnJkOwLZPM3DYoff/hdP7nXBycwy2f5ZxuVXdkMLCgnn+xQ6MHrmYPo9MJy7uHE882dK0TJmy4fz7qVa8OPRj+j0+iwUffc0rYx4EoNMD9QgrFswTvWbSt8/7VKtWhmYtqhZEKZIPvLZHHz9+nISEBAA+++wzxowZw+rVq721+t+1facfVas4KFfG+Uf9gU65rN3gj3HJ3/iDh600qOsgKsI5s0UTO1/H+mGzwbMDbQzqbwMgKdlCjs1CkSK+00GI3WmheoxB+TLO6X91dLD6C6up/gOHLDSsaxAd6Zxu1cRgU6wFmw2+32/BaoXHn/Hjwcf8mT7Xit3u/TquVexOK9Vi8rZ/l4521nzhZ6r/x0MW7qjrIMpVv4PNsVZszs3Ozj1Wtu6w8lBHHyr8op3f+lGlioMyF+vv1NHGhg0BpvoPH7ZSt66diIvv/yZNcomNdaYuNWvaefSRbKxW8PODipUcJCT4Rmfhn77tj+/JIbpSADeVcnYKa7UtzIFNFzAueQEMh+FMTjKd83IuGPgFOBOiMtUCaNQ1GIvVgtXPQuQt/qQmOrxfyDWq2+BmDh+M49TJcwCsXLabVq2rmZax2exMfPtzks86O4WHD8ZxU3hR/P2tLP50B2NeXophQGhoMEVCCpGWesHrdUj+uOohttmzZ/Puu+9it9spVaoUVapUcd0qV65MmTJlrrqSOXPmMG/ePBwOBw0bNiQuLo7WrVuzePFijh07xtNPP31dirlWCYkWoiLzdurICIOMDAsZmbii9moxDj5d4k9cvIWS0Qar1vpjs1lISYUSxcHfD15+PZCNm/xo1sRO+bK+00GKP2MhOiKvvVERkO5Wf/WqBguXWDkdD6WiYfkaKzabhfOpkGuHRvUcPNvPwYVsGPCiH0WCrTzaxTc+KBPOWFwdX3Buf/f6q1V18PES/0vq93PVjwHj3vFn6rgcFq/wvaPWiWesRF7y/o+4+P7PzMR1mC0mxsGSpYHEx1uIjjZYuzYAm81CaqqF+vXzOgbx8RYWLw7guWd944/EP33bpyXaCSmR15kNKWElJ9MgJ8sgKNjZCQosbKX10yEsHHqOQqFWDIdB97dvAuDmOnlJY8oZO7tWZNHm6RDvFvEXREaGcuZMqms6MTGVIkULERwc6DrMlhCfQkJ8imuZfgPuJvabw+TmOvcZu93BE0+2oFPnehw+FMe+vSe8W0Q+0Vlsf6CDNGPGDMaOHUvNmjX59ddfOXz4MIcOHWLz5s389NNPAFSqVIlFixZd8TkWL17M6tWrSUpKokOHDmzbto2goCC6dOnCQw89VOAdJOMKf8f9LvkSXPt2B0/0tDFsVBAWi8F9be2EhhoEXPIKjh6Rw7BnYfioID74KIB/97Hlb8OvkysdLrdeUn+92w369euvPfMAACAASURBVLLzzEh/rBaD+9sZhF2s/6EOBr+N6AsMhEe7OFi4xPr/7d13dFRl/sfxz6QnklBMobpUF9ClSlMgFGkJSFURFaSjAhKV5iIgCEvToIAsC4qiIAtKMZYISlmFqIAKQRGkLiFAQkASCJA2vz/iDt4bIJPd5E7ym/frnDknd/LM3Od77gS+87nPndHjDxb93AvDzer/4/FvXN+uYQOy9PyLPrLZpO4R2SodZJeHTRr3ko+eH5mlkNutmW9hc+b416+frf79MzR5sr9sHlKXLpkKCrLLy+v6gw8e8tDkyf7q0SNTLVqUjDTF3Y+9/Sb12zyuryFKPp6luNXpGvhGOZWt4KU9H6Vr499SNeD1srLZcsedOZypDTMuqlGkv2o0LSGLzyTH/M1utIbIz89bYyd2U2hokCaMNf5/t2zJVi1ftl3PjovQM8910ZyZMUUyX1gr3wapVKlSatOmjby8vBQaGqrGjRsbfp+QkOBolG4mJydHPj4+qlSpkgYNGiTfP6zezC4G52LCwuzaf+D6v4jJyTYFBdrl7399zOV0qWGDHD0QmfvOOOW8tGS5t4KCpG++81CN6naFBNsV4C91bJ+lrf8qOe8mK4TaFf+H+pPOSUGBubX8x+X03CapV2TuAtSU89KitzxUOkiK2WTTn2vYdWeN3LF2e26iVlKUDzUe///Ubz7+jernqEdk7us15by0+C0vJZy2KfG0Ta8u8vr9fpuyc6SMDGnyuJKxWDcsNEcHDlx/vSYn2xRoqj89XWpQP0uREblN//nzNi1f7qug3xcqb9nipfmv+Wn06Ku6v33JqFvi2AeGeOr0oetv5NJScuRXyiYfv+uNw7HvM1SpjrfKVsits2Gkv7a+eUlXUu0KKG3TgX9d1ReL09R+eKDqtvGzvIb/RdLZVNWpW8mxHRwcqNTUK7p61fjmNjQ0SNNnPaR/nzin5555TxkZucf3rrsr67ff0nUq4byys3O06bN9Gjmmk6U1FBm+aiT/NUjDhg3T2rVrb/r7ypUrq23btrd8jo4dO+qxxx5Tdna2Ro0aJUn65Zdf1K9fP3Xp0qWAUy58ze7J1v4Dnvp3Qu4LYl2Ml1rdZ2zczp2z6ckxvrr0+/rDt971Vsd2WbLZpC+2eWnZO7lrljIycrfvaej6xs9ZLZrYte9nm04k5G6v/chDbe8zvoNKOicNGuPlqH/JCg91aZcjm006fMymRW95KjtbunpNWr3eQ53alYzTa5LUokmO4n/2cBz/Dz/yUrjp+Cefs2nYGB9H/UtXeKlTu2zVv8uuz9Ze0+o3cxfw9n4gWx3bZpeY/yAl6Z57snXggKcSfq8/JsZb991rnP+5czaNiQrQ5d/rf/ddH7VrmymbTdq+3UsLFvpq7pz0EtUcSRz7qg19lHgwUxcSc+e897MrqtnMmACF1fDSyf0Zunwh92/612+uqXSYpwJKe+jgjqva8o80PTitTIlrjiRpz66jqlO3oipVzj1l2K17I+38+pBhTGCgn15Z8Li+/tdBzXhpg6M5kqQGjarqqVEd5OFpk82We7XbD98ft7IEFCGb3X6zkDVXw4YNlZmZqdatW6tVq1aqU6eO/vznP8v/j2+xnLBr1y41adLEsX306FGdPHlS4eHhBZ70b4lVCvyY/Oz4xkNvLPVRVpZUqaJdUyZeU+JpD82Y66P3luWmRmvXe+mDDV7KsUv1787R889kyM9XSrskzXrVR0ePechmk1q3zNawJzINpygKi7/NO/9B/4Wvvsm9zD8zU6pS0a4ZL2QrIdGmqXM9tfbN3H8Q3l/nodUbcj/KoNFf7Jr4TLb8fKUrV6W/veapfT/blJUldWiTo9FDcpunwpalomk8v/7m+qXelSvaNf2FTJ1KtGnaXG+tfjN3LcLqdZ5asyF3AW+Dv+Ro/DNZ8jOdTfj7ci/9dlFFdqn3xZyiOW37zTeeWrrMV1lZUsWKdk2ccEWnT3to7jw/LVuaLklav95bGzZ6y55j091/ydIzo6/J11d67PHbdOmSFBx8/Z+Su+/O1phnrhXqHEt7FM1rv6Qc+/fT7iiS5z26+5r+9c5lZWfZVaa8pyKeDdLFM9mKXZCmJ14vJ0n6/pN0/fDxFXl62eQXaNP9wwMV/CcvLR2WomuXc1Tq9uuRcaU63urwZOGvQ1o9pHOhP6ckNW1eQ4OHtZWXt6dOn7qg2TM+UoWKZfXsuEiNGLxM/R6/TwMGtdaxo8mGx42LWqn09Gt6alRH1Wtwh+x2u/bvO6m/L/pC164V/mvgi3/9tdCf81aqz3/Vsn0dHfOsZfsqiHwbpJMnT+qXX37RwYMHdfDgQf3yyy9KTExU5cqV9fnnn1s1T4OiaJBKiqJqkEqKomqQSoqiapBKgqJqkEqKomqQSoqiapBKChok6+W7UKZKlSqqUqWKOnTo4LgvPT1dhw4dusWjAABAicVVbP/d5yAFBASoQYMGhT0XAACAYqHkXGoFAAAswecg8VUjAAAAeZAgAQAAIxIkEiQAAAAzGiQAAAATTrEBAAAjTrGRIAEAAJiRIAEAAAMu8ydBAgAAyIMECQAAGNmL4NvGSxgSJAAAABMSJAAAYMQaJBIkAAAAMxIkAABgwFVsJEgAAAB5kCABAAAjEiQSJAAAADMSJAAAYMAaJBIkAACAPEiQAACAEQkSCRIAAIAZDRIAAIAJp9gAAIARp9hIkAAAAMxIkAAAgAGX+ZMgAQAA5EGDBAAAYEKDBAAAYMIaJAAAYMQaJBIkAAAAMxIkAABgwFVsJEgAAKCEiImJUUREhDp06KCVK1fedNy2bdvUrl07x/auXbvUrFkzde/eXd27d9fEiRPz3VeJTJD61r7f1VNwGZuvr6un4FLZF1NdPQWX8iwd5OopuExOWpqrp+BSORkZrp6CS3n6/OzqKbiXYpggnT17VtHR0Vq3bp18fHzUt29fNWvWTDVr1jSMO3funGbPnm24Lz4+XoMGDdLw4cOd3h8JEgAAKPZ27typ5s2bq0yZMgoICFCnTp0UGxubZ9ykSZM0cuRIw33x8fHasWOHevTooREjRuj06dP57q9EJkgAAKAIWZggpaamKjU179mBoKAgBQVdT82TkpIUEhLi2A4NDdW+ffsMj1mxYoXq1q2r+vXrG+4PDAxUZGSk7r//fr3//vuKiorS6tWrbzkvGiQAAOAy77zzjhYuXJjn/pEjR2rUqFGObbs9b9dms9kcPx86dEibNm3S22+/rTNnzhjGTZs2zfHzI488oldeeUVpaWkKDAy86bxokAAAgIGVV7ENGDBAPXv2zHP/H9MjSQoLC9Pu3bsd20lJSQoNDXVsx8bGKjk5Wb1791ZmZqaSkpLUr18/vffee1qyZImGDRsmT09Px3gvr1u3QDRIAADAZcyn0m7m3nvv1YIFC3T+/Hn5+/tr06ZNmj59uuP3o0eP1ujRoyVJCQkJ6t+/v1atWiVJ2rx5s/70pz8pIiJCGzZsUP369eXv73/L/bFIGwAAFHthYWGKiopS//791aNHD3Xt2lX16tXT0KFDFR8ff8vHzp49WytWrFBkZKQ+/PBDvfzyy/nuz2a/0Um9Yq5z0EBXT8FluMyfy/zdFZf5u/dl/h4+Pq6egkt9fvXmn/lTFOq8GG3Zvg5Mj7JsXwVBggQAAGDCGiQAAGDAV42QIAEAAORBggQAAIxIkEiQAAAAzEiQAACAEQkSCRIAAIAZCRIAADDgKjYSJAAAgDxIkAAAgBEJEgkSAACAGQkSAAAwIkEiQQIAADAjQQIAAAZcxUaCBAAAkAcNEgAAgAmn2AAAgBGn2EiQAAAAzEiQAACAAYu0SZAAAADyIEECAABGJEgkSAAAAGYkSAAAwIgEiQQJAADAjAQJAAAY2Fw9gWKABAkAAMCEBAkAABixBokG6T+adqqngVP6yNvXS8f2Jyh65FtKT7t6w7HPLR6s4z+f0ocLYg33B1cqp/lfTtJT905W6vlLVky70DTpcLcG/rVHbv0/n9L8Z95V+qUb1//sggE6cSBRH76xWZLk4+etp2f3Va0GVeXhYdPB749p0fjVyriaaWUJBda0SwMNerlvbs3xJ/XqsH8oPe2KU2M8PGx6+rWBqte6tiTpu89+1NIJqyRJ9cPraticR+Xp5anUlDT9/fl3dXTfvy2vryDc7fg37dxAA6c/7Diu0SOW5j32NxkTWPY2jXp9kKrXv0NXL1/TphX/0keLN0mS7mxcXSPmPS6/AF95eHpozSsx2vL+DleUeEtNIxpp8Mx+8vb11rF9J/TKkMV567/JmBfXPKdKNcs7xpWvFqp923/W5B6zdUedyopaMlz+pfxkt9v15sSV2r1pr9Xl5cvdjz+cwyk2SaVvD9SzbwzW9McXaUjjF3T6eLIGvvRgnnFV7qygWTHj1Kpnkzy/a//IvXoldqKCK5a1YsqFqvTtpfTsa/318qB/aGiLqTpz/JwGvtgzz7gqtcrrb+vGqNUDjQ33943qIg9PTz3d5mU9FT5dPn4+eviZzlZN/79SOjhQzy8drmkPz9fgu5/X6WNnNXhGX6fHtH+0larcWUHDG47XiMYTVa91HbXq3UwBQf6avCZKSyes0ojGE7Rg1HL9ddVoefsU3/ci7nb8SwcH6rl/DNP0vvM1pN5YnTmWpEEvP+z0mOFzH9OVy1c1rME4jWk9RU061VezLg0lSS+ufkbvTv9ATzV7QZO6z9Hw2Y+qYo0wy2u8ldLBQXr+rac0rc88DarzTO7retajTo+Z/tArGtForEY0GqtXh/1dl367rAUjl0mSRi8aotjlWzSi0VjNG/yGJv3zWXl4Fq//Ztz9+DvLZrfuVly55JU7a9YsV+z2phq1v0uHvj+mxCNnJUmfvLlF7R5snmdct2HttXnlV/pq/S7D/eXKl9G9kY30Yp9oS+Zb2Bq1qatDP55Q4tEkSdLHb/9Lbfs0zTOu6+A22vx+nL76aI/h/v1xv2r1q5/KbrcrJ8euI/EnFVqlnCVz/2817lBPB3cfVeLhM5Kkj5d8oXaP3Of0GE9PD/nd5itvX295+3rJy8dLmVczValmeV2+mK4ft/4kSTp5MFHpqVdUp3ktC6srGHc7/o3u/4sO7jnq+Hv/eOkXatf3PqfH1GpYTV+u+lo5OXZlZWbru89+VMteTeXt6633ZqzTD1tyj/25U+d1MeWSQirfbmF1+WvcsZ4O7TqiU7+/rmMWb1L7fq0KPMbL20vj3h6pxVFvKzkhRZLk4emhwLK3SZICAv2VcTWjqMspMHc//nBekb+tnThxYp77tmzZoosXL0qS/va3vxX1FPIVUqmckhPOO7aTT13QbaUDFBDoZzjN9sbz70mSGoTXNTz+/JnfNP2xhdZMtggEVyqr5FMXHNvnEi/otiB/BZTyM5xmWTxhtSSpwe+nlf7j+20HHD+HVi6nHsPb6fVnVxbxrP83IZXLOf5Rl6TkhPO/H3N/R9R+qzGbVmxXq97NtOr4Inl6eWjPF/H65pPvFRDoL/9Sfmp8/1+054t43dm4uv5Ut7LKVShjeY3OcrfjH1L5dp3749/7DY/9zcf8suuI2vdrqZ92HpK3r5da9myirMxsZV7L1Odvb3c8psvgtvIv5asD3/5qXXFOCKkSrOSEc47t5ISUvPU7Mabz4HZKSTyvHRu+c4xbMHKZ5n45Rb3GdFWZ0NKa+Ui0crJzLKrMOe5+/J1WjJMdqxR5glSmTBlt27ZNtWvXVtOmTdW0aVMFBAQ4fi4ObB43vqAxu5j9YRcVj5vVn1Ow+mvWu0NzY55XzJvb9N3m+MKYWpGxedz4pf/Hf8xvNeaxF3vr4rlUPVx5hPpVG6nAsrep95gIpadd0ZTer6jv+O5avPtvuv+xVvpx60/KysgqkjoKg7sd/5vW+4djf6sx/xi/Una7XW98O0NT1kTp+y/35zm+Dz3fTY9P6q0pvV4pdmuxblZbjhP1/3FM7zGRWjnjQ8e2t6+3Jq2O0tyBi9TvjhF6Lnyynvn78GKXoLj78YfzirxBGj9+vF599VV9+umnqlixonr27KnSpUurZ8+e6tkz7zoHV0hOOK9y5a+/ww+uWFZpFy7pWnrxi4eLQlLCeZULK+3YDq5QRmkXLheo/vAe92jmB89o+fT1+uf82Pwf4GLJJ88ZUp3gSuWUev6SrqZfc2pMyx5N9Pnb25WVma301Cva/O5Xqh9eVzabTVcvX9XYDi/ryXsm6o2od1SxRpgjqi+O3O34J51MMf69VyqntPOXdO0Px/5WYwKC/PXmC+9reOMJmhg5S/acHMfx9fbx0oQVT6vtQy0U1WaqjsYXv8X5Sf8+p3Llr6+VvNFrP78xNRpUlaeXp/Zt/9kxptrdVeQb4KtvP/leknTg21914qeTqt2seJ1edvfjD+dZsgapRYsWWrJkiVatWqXZs2crOzvbit06bc+X+1W7SXXHYrrIQW0V98kPLp6Vdb7fdkC1G1dTxeqhkqSIJ1orLtb5K09admukETMf0l8ffE3b1u3K/wHFwJ7N8arTtJYq/n41Ttdh7RUXs8fpMb/+cFyt++SuU/P08lSLbo30y3eHZbfb9fLGcarVqJokqVXvZsrKzC7WV7G52/Hf80W8ajetef3vfWh7xX28x+kxXYe2V//JfSRJZUKD1GVQW239505J0l9XjVZAoL/GtHlJZ0+cU3G0Z9Ne1Wley3ElWtcRHRW3cVeBxtQLr6sft+43PObU4TO6rXSA6ra4U5JUoXqY7qhTSYd/OFaU5RSYux9/p9ktvBVTNrvdbun01q5dq88++0xvvfXWf/0cnYMGFuKMcjXpWE8Dp/SWl4+XTh9L0tzhy1ShaojGLBiop1tOMYy92WX+khSbulwPVR1VZJf523x9i+R5m9x/t574aw95+Xjq9PFkzXv6bVX4U7Cemf+4RradYRhrvsx72bfTVCrIX+fO/OYY8/N3R/TG+NWFPs/si6mF9lxNOjfQoJcflrePlxKPnNXcQYtVvlqonl0yVE82eeGmY9IuXFZguVJ6ev4A1WpYTdnZOfpxy34tGbdS2VnZ+kur2nrylf7y8vHS+dO/af5Ty3TmWFKhzNmzdFChPI9ZSTj+OWlphfZcTTrV16DpD+f+vR9N0tzBucc+avFQPdXshZuOSbtwWf6l/DTurSdVsUaYbDabVs/9SFve36G6Le5U9NYpOnkoURlXrp9WefOv72vPF//7KcecjMJLtJt2aahBM/s5XtdzBixUheqhenbpkxrRaOxNx6RdyP13bdTCwUo5fUGrZqwzPG/9Nndp6OzH5OPno6zMLL03/QPt3Fg4TbOHj0+hPI9UMo//51etXddXf7R1Fx3tfT3Ksn0VhOUNUmEoigappCiqBqmkKMwGqSQqqgapJCjMBqkkKswGqSQqzAapJLK6QWowyroG6ccFxbNBKl4fUAEAAFAMFN9PrwMAAK5R4s4tFT4SJAAAABMSJAAAYFCcvwLEKiRIAAAAJiRIAADAiASJBAkAAMCMBAkAABiwBokECQAAIA8SJAAAYESCRIIEAABgRoIEAACMSJBIkAAAAMxokAAAAEw4xQYAAAy4zJ8ECQAAIA8SJAAAYESCRIIEAABgRoIEAAAMbHYiJBIkAAAAExIkAABgRIBEggQAAGBGggQAAAz4HCQSJAAAgDxIkAAAgBEJEgkSAACAGQkSAAAwYA0SCRIAAEAeJEgAAMCIBIkECQAAwIwGCQAAwIRTbAAAwIBF2iRIAAAAeZTIBCn70iVXT8F13Ll2KCslxdVTAFwiJyPD1VNwLyRIJEgAAABmJTJBAgAARYc1SCRIAAAAeZAgAQAAIzsREgkSAACACQkSAAAwYA0SCRIAAEAeJEgAAMCIBIkECQAAwIwECQAAGNhyXD0D1yNBAgAAMCFBAgAARqxBIkECAAAwo0ECAAAwoUECAAAGNrt1t4KIiYlRRESEOnTooJUrV+b5/ebNm9WtWzdFRkZqwoQJysjIkCQlJibq0UcfVefOnfXkk0/q8uXL+e6LBgkAABR7Z8+eVXR0tFatWqWNGzfqn//8pw4fPuz4fXp6uqZNm6bly5frk08+0bVr17R+/XpJ0ksvvaR+/fopNjZWd999t954441890eDBAAAjOx2625O2rlzp5o3b64yZcooICBAnTp1UmxsrOP3AQEB2rJli4KDg5Wenq6UlBQFBQUpMzNTu3btUqdOnSRJvXr1MjzuZriKDQAAuExqaqpSU1Pz3B8UFKSgoCDHdlJSkkJCQhzboaGh2rdvn+Ex3t7e2r59u8aNG6fQ0FC1bNlSFy5cUKlSpeTlldvyhISE6OzZs/nOiwQJAAAYWLkG6Z133lH79u3z3N555x3DnOw3SJtsNlue+8LDw/Xtt9+qbdu2mjp1qtOPMyNBAgAALjNgwAD17Nkzz/1/TI8kKSwsTLt373ZsJyUlKTQ01LH922+/af/+/WrZsqUkqVu3boqKilK5cuV06dIlZWdny9PTU8nJyYbH3QwJEgAAMLJbdwsKClLlypXz3MwN0r333qu4uDidP39eV65c0aZNm9S6devrU7bbNXbsWCUmJkqSPvvsMzVq1Eje3t6655579Omnn0qSNmzYYHjczdAgAQCAYi8sLExRUVHq37+/evTooa5du6pevXoaOnSo4uPjVbZsWU2fPl3Dhw/XAw88oOPHj2vs2LGSpClTpmjNmjWKiIjQ7t27NWbMmHz3Z7Pf6ORcMdfB40FXTwEAAMtszllr6f5a9Zxn2b6+Wv+8ZfsqCBIkAAAAExZpAwAAo5J3cqnQkSABAACYkCABAACDgn5H2v9HJEgAAAAmJEgAAMCIBIkECQAAwIwGCQAAwIRTbAAAwIBF2iRIAAAAeZAgAQAAoxwiJLdukJpGNNLgmf3k7eutY/tO6JUhi5WedsWpMS+ueU6VapZ3jCtfLVT7tv+syT1m6446lRW1ZLj8S/nJbrfrzYkrtXvTXqvLyxf1u2/97ly7RP3U7971wzlu+2W1pYODtHT/q4pqOUmnDp/RkFmPyj/QXwueXlagMZJ05z01NHntc4pq9aKSE1I0b8tUbX53uz5fvlU1GlTVK1tfUq/ggcrJzvmf511YqN9963fn2iXqp/6SWb/VX1YbHjHHsn1t/3ScZfsqCEvWIO3bt8/xc1xcnGbNmqV58+Zp717XddaNO9bToV1HdOrwGUlSzOJNat+vVYHHeHl7adzbI7U46m0lJ6RIkjw8PRRY9jZJUkCgvzKuZhR1OQVG/e5bvzvXLlE/9bt3/XCeJafYpkyZovXr12vlypVavXq1evfuLUmaPHmyHnzwQT322GNWTMMgpEqwkhPOObaTE1J0W+kABQT6O6JWZ8Z0HtxOKYnntWPDd45xC0Yu09wvp6jXmK4qE1paMx+JLlbvoCTqd+f63bl2ifqp373rdxZXsVm8BmnNmjVasWKFypYtK0nq06eP+vTp45IGycPDdsP7//hidmZM7zGRih6+xLHt7eutSaujNHfgIn37yfeq06yWpn00QQd3HXG8yygOqN9963fn2iXqp373rh/Os+QUW1ZWlnJycnT77bcrICDAcb+Pj488PFzzSQNJ/z6ncuXLOraDK5VT6vlLupp+zekxNRpUlaeXp/Zt/9kxptrdVeQb4KtvP/leknTg21914qeTqt2sVlGXVCDU7771u3PtEvVTv3vX7zS73bpbMWVJd1K2bFmFh4fr8OHDmjJliqTctUh9+/ZV586drZhCHns27VWd5rUcVyN0HdFRcRt3FWhMvfC6+nHrfsNjTh0+o9tKB6huizslSRWqh+mOOpV0+IdjRVlOgVG/+9bvzrVL1E/97l0/nGfJKbYVK1ZIko4eParU1FRJuenR6NGj1aZNGyumkMdvyamaN+gNvbj2OXn7eCnxyFnNGbBQdzaurmeXPqkRjcbedMx/VK5VQWeOJxme9/LFdE3tNVdPzR8oHz8fZWVmaf6If+j00bNWl3hL1O++9btz7RL1U7971+8s1iC58WX+AACUFFZf5t+202zL9rX18/GW7asg3PqDIgEAwA2UuOik8PFdbAAAACYkSAAAwMBW8lbfFDoSJAAAABMaJAAAABNOsQEAAKOS+Q0phYoECQAAwIQECQAAGLBImwQJAAAgDxIkAABgRIBEggQAAGBGggQAAIxYg0SCBAAAYEaCBAAADGwESCRIAAAAZiRIAADAiDVIJEgAAABmJEgAAMDAxnexkSABAACYkSABAAAj1iCRIAEAAJiRIAEAACMCJBIkAAAAMxokAAAAE06xAQAAAxuLtEmQAAAAzEiQAACAEQkSCRIAAIAZCRIAADDiq0ZIkAAAAMxIkAAAgAFXsZEgAQAA5EGCBAAAjEiQSJAAAADMSJAAAIARCRIJEgAAgBkJEgAAMOJzkEiQAAAAzEiQAACAAZ+DRIIEAACQBw0SAACACafYAACAEafYSJAAAADMSJAAAIARCRIJEgAAgBkJEgAAMCJBIkECAAAwI0ECAABGfNUICRIAAIAZCRIAADDgq0ZIkAAAAPIgQQIAAEYkSCRIAAAAZiRIAADAKIcEiQQJAADAhAQJAAAYsQaJBAkAAMCMBgkAAMCEU2wAAMCIU2wkSAAAAGYkSAAAwIgEiQQJAADAjAQJAAAY8UGRJEgAAABmNEgAAMDInmPdrQBiYmIUERGhDh06aOXKlTcdN378eK1bt86xvWHDBrVs2VLdu3dX9+7dFR0dne++OMUGAACKvbNnzyo6Olrr1q2Tj4+P+vbtq2bNmqlmzZqGMVOmTFFcXJyaNWvmuD8+Pl4TJkxQ165dnd4fCRIAADCy2627OWnnzp1q3ry5ypQpo4CAAHXq1EmxsbGGMTExMWrfvr26dOliuD8+Pl4bNmzQAw88oOeff14XL17Md380SAAAwGVSU1OVkJCQ55aammoYl5SUpJCQEMd2aGiozp49axgztbnNkwAAC5pJREFUZMgQPfjgg3n2ERISolGjRmnjxo2qUKGCpk2blu+8OMUGAACMLLyK7Z133tHChQvz3D9y5EiNGjXKsW2/Qdpks9mc2seiRYscPw8ZMkT3339/vo9x6wapaUQjDZ7ZT96+3jq274ReGbJY6WlXnBrz4prnVKlmece48tVCtW/7z5rcY7buqFNZUUuGy7+Un+x2u96cuFK7N+21urx8Ub/71u/OtUvUT/3uXX9xM2DAAPXs2TPP/UFBQYbtsLAw7d6927GdlJSk0NDQfJ8/LS1NH374oZ544glJuY2Wl1f+7Y/NfqOWrJjr4JE3Piuo0sFBWrr/VUW1nKRTh89oyKxH5R/orwVPLyvQGEm6854amrz2OUW1elHJCSmat2WqNr+7XZ8v36oaDarqla0vqVfwQOVkF2y1flGifvet351rl6if+ktm/Ztz1v7Pz1EQXao8Y9m+Pjv5mlPjzp49q0ceeUQffPCB/P391bdvX02fPl316tXLM3bChAlq2rSpevXqpezsbIWHh2vRokWqX7++Fi5cqKSkpHxPs1m2Bumrr75ynE/csGGDpk2bpg8//NCq3efRuGM9Hdp1RKcOn5EkxSzepPb9WhV4jJe3l8a9PVKLo95WckKKJMnD00OBZW+TJAUE+ivjakZRl1Ng1O++9btz7RL1U79711+ShYWFKSoqSv3791ePHj3UtWtX1atXT0OHDlV8fPxNH+fp6an58+dr6tSp6tKli3766SeNHTs23/1ZcoptxowZOnDggKKjozV//nzFx8erffv22rx5sw4cOKBJkyZZMQ2DkCrBSk4459hOTkjRbaUDFBDo74hanRnTeXA7pSSe144N3znGLRi5THO/nKJeY7qqTGhpzXwkuli9g5Ko353rd+faJeqnfveu32nF9ORSt27d1K1bN8N9S5cuzTNu1qxZhu177rlH69evL9C+LGmQduzYoZiYGHl6emrbtm1as2aNfHx89PDDDxfoMwkKk4fHjRd2/fHF7MyY3mMiFT18iWPb29dbk1ZHae7ARfr2k+9Vp1ktTftogg7uOuJ4l1EcUL/71u/OtUvUT/3uXT+cZ8kpNj8/P6Wk5L5Abr/9dqWnp0uSrly54tRCqaKQ9O9zKle+rGM7uFI5pZ6/pKvp15weU6NBVXl6eWrf9p8dY6rdXUW+Ab769pPvJUkHvv1VJ346qdrNahV1SQVC/e5bvzvXLlE/9bt3/XCeJQ3SyJEj1adPH82ePVvVq1fX448/rpkzZ+qhhx7SwIEDrZhCHns27VWd5rUcVyN0HdFRcRt3FWhMvfC6+nHrfsNjTh0+o9tKB6huizslSRWqh+mOOpV0+IdjRVlOgVG/+9bvzrVL1E/97l2/04rhB0VazbKr2E6ePKkvvvhCJ06cUHZ2toKDg9W2bdsbrj7PT2FcxSZJTbs01KCZ/eTt46XEI2c1Z8BCVageqmeXPqkRjcbedEzahUuSpFELByvl9AWtmrHO8Lz129ylobMfk4+fj7Iys/Te9A+00/QHWBxQv/vW7861S9RP/SWvfsuvYqs0Kv9BheSzUwss21dBuO1l/gAAlBSWN0gVnrZsX5+dXpT/IBfgq0YAAABM3PqTtAEAwA2UvJNLhY4ECQAAwIQECQAAGJEgkSABAACYkSABAACjHBIkEiQAAAATEiQAAGBgt5fQL9ktRCRIAAAAJiRIAADAiDVIJEgAAABmJEgAAMCIz0EiQQIAADCjQQIAADDhFBsAADDK4TJ/EiQAAAATEiQAAGDEIm0SJAAAADMSJAAAYGBnDRIJEgAAgBkJEgAAMGINEgkSAACAGQkSAAAw4stqSZAAAADMSJAAAICRnavYSJAAAABMSJAAAICBnTVIJEgAAABmJEgAAMCINUgkSAAAAGY0SAAAACacYgMAAAYs0iZBAgAAyIMECQAAGLFIWza7na/sBQAA+CNOsQEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIAEAAJjQIBVATEyMIiIi1KFDB61cudLV07HcpUuX1LVrVyUkJLh6KpZbuHChIiMjFRkZqTlz5rh6OpZ77bXXFBERocjISC1fvtzV03GZ2bNna8KECa6ehuX69++vyMhIde/eXd27d9fevXtdPSVLbdmyRb169VLnzp318ssvu3o6sAjfxeaks2fPKjo6WuvWrZOPj4/69u2rZs2aqWbNmq6emiX27t2rSZMm6fjx466eiuV27typr7/+WuvXr5fNZtOQIUO0efNmdejQwdVTs8R3332nb775Rh999JGysrIUERGh8PBwVa9e3dVTs1RcXJzWr1+vNm3auHoqlrLb7Tp69Ki2bdsmLy/3+y/j5MmTmjJlitauXavbb79dAwYM0Pbt2xUeHu7qqaGIkSA5aefOnWrevLnKlCmjgIAAderUSbGxsa6elmXWrFmjKVOmKDQ01NVTsVxISIgmTJggHx8feXt7q0aNGkpMTHT1tCzTtGlTrVixQl5eXkpJSVF2drYCAgJcPS1L/fbbb4qOjtaIESNcPRXLHT16VDabTUOHDtUDDzyg9957z9VTstTmzZsVERGh8uXLy9vbW9HR0apfv76rpwULuN/bgf9SUlKSQkJCHNuhoaHat2+fC2dkrRkzZrh6Ci5Tq1Ytx8/Hjx/Xp59+qtWrV7twRtbz9vbW66+/rrfeekudO3dWWFiYq6dkqcmTJysqKkqnT5929VQsl5qaqhYtWmjq1Km6evWq+vfvr2rVqum+++5z9dQsceLECXl7e2vw4MFKTk5W27ZtNWbMGFdPCxYgQXKS3W7Pc5/NZnPBTOAqv/76qwYNGqTx48eratWqrp6O5UaPHq24uDidPn1aa9ascfV0LLN27VpVqFBBLVq0cPVUXKJhw4aaM2eOAgICVK5cOfXp00fbt2939bQsk52drbi4OM2dO1dr1qxRfHy81q9f7+ppwQI0SE4KCwvTuXPnHNtJSUluebrJXe3Zs0dPPPGEnnvuOfXs2dPV07HUkSNHdODAAUmSv7+/OnbsqIMHD7p4Vtb59NNPtWPHDnXv3l2vv/66tmzZopkzZ7p6WpbZvXu34uLiHNt2u92t1iIFBwerRYsWKleunPz8/NS+fXu3OnvgzmiQnHTvvfcqLi5O58+f15UrV7Rp0ya1bt3a1dOCBU6fPq2nn35a8+bNU2RkpKunY7mEhARNmjRJGRkZysjI0JdffqnGjRu7elqWWb58uT7++GNt3LhRo0ePVrt27fTCCy+4elqWSUtL05w5c3Tt2jVdunRJ69evd5sLFCSpbdu2+vrrr5Wamqrs7Gx99dVXuuuuu1w9LVjAfd4G/I/CwsIUFRWl/v37KzMzU3369FG9evVcPS1Y4M0339S1a9c0a9Ysx319+/bVI4884sJZWSc8PFx79+5Vjx495OnpqY4dO7plo+iu2rZt6zj+OTk56tevnxo2bOjqaVmmfv36GjJkiPr166fMzEzdd9996t27t6unBQvY7DdaXAMAAODGOMUGAABgQoMEAABgQoMEAABgQoMEAABgQoMEAABgQoMEAABgQoMEAABgQoMEwCldunRR69at9euvv7p6KgBQ5GiQADjl448/VtWqVfX555+7eioAUORokAA4xdPTU40bN3arL6oF4L74LjYATrl69ao++eQT8e1EANwBCRIAp0RHRyssLEwnT57U5cuXXT0dAChSNEgA8vXDDz8oNjZWCxYsUGBgoA4dOuTqKQFAkaJBAnBL165d08SJE/XSSy+pTJkyql27NuuQAPy/R4ME4JZee+01NWzYUG3atJEk1a5dW7/88otrJwUARYwGCcBN7du3T7GxsXrhhRcc99WpU4cECcD/ezY7l6QAAAAYkCABAACY0CABAACY0CABAACY0CABAACY0CABAACY0CABAACY0CABAACY0CABAACY/B/XbCAzaIVlFwAAAABJRU5ErkJggg==\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "# visual representation of grid search\n",
     "# uses seaborn heatmap, you can also do this with matplotlib imshow\n",
@@ -2679,513 +2439,10 @@
   {
    "cell_type": "code",
    "execution_count": 11,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.18333333333333332\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.18611111111111112\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.13055555555555556\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.24444444444444444\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.23333333333333334\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.12777777777777777\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  1e-05\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.1527777777777778\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.9111111111111111\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.8888888888888888\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.8722222222222222\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.8305555555555556\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.8888888888888888\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.8805555555555555\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.0001\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.8944444444444445\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.975\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.9777777777777777\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.9805555555555555\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.9861111111111112\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.9805555555555555\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.9777777777777777\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.001\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.9444444444444444\n",
-      "\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/usr/local/lib/python3.7/site-packages/sklearn/neural_network/multilayer_perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (100) reached and the optimization hasn't converged yet.\n",
-      "  % self.max_iter, ConvergenceWarning)\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Learning rate  =  0.01\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.9861111111111112\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.9888888888888889\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.9888888888888889\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.9861111111111112\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.9888888888888889\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.9722222222222222\n",
-      "\n",
-      "Learning rate  =  0.01\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.9527777777777777\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.925\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.9194444444444444\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.9\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.8472222222222222\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.8666666666666667\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.9138888888888889\n",
-      "\n",
-      "Learning rate  =  0.1\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.775\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.16944444444444445\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.16111111111111112\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.18611111111111112\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.18333333333333332\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.10555555555555556\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.11388888888888889\n",
-      "\n",
-      "Learning rate  =  1.0\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.11388888888888889\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  1e-05\n",
-      "Accuracy score on test set:  0.1111111111111111\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.0001\n",
-      "Accuracy score on test set:  0.09166666666666666\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.001\n",
-      "Accuracy score on test set:  0.1361111111111111\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.01\n",
-      "Accuracy score on test set:  0.11388888888888889\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  0.1\n",
-      "Accuracy score on test set:  0.10555555555555556\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  1.0\n",
-      "Accuracy score on test set:  0.10555555555555556\n",
-      "\n",
-      "Learning rate  =  10.0\n",
-      "Lambda =  10.0\n",
-      "Accuracy score on test set:  0.058333333333333334\n",
-      "\n"
-     ]
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "from sklearn.neural_network import MLPClassifier\n",
     "# store models for later use\n",
@@ -3215,29 +2472,10 @@
   {
    "cell_type": "code",
    "execution_count": 12,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAkIAAAJgCAYAAABmwww9AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4wLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvqOYd8AAAIABJREFUeJzs3XmcjeX/x/HXfc6ZFaMwYyfKUoyyJEsqStaI+IZCRF8xRKX6fi0pirL/olDIXmSnUBT6WooWsrZQCTPGMMOsZ7l/fxyNuRlp4Rxnzvv5eMyD+z7Xue/rc865r/s6n+u672OYpmkiIiIiEoRs/q6AiIiIiL+oIyQiIiJBSx0hERERCVrqCImIiEjQUkdIREREgpY6QiIiIhK01BES8aERI0bQunVrWrduTdWqVWnSpEn2ckZGxp/ezvr16xkxYsQflomPj6dDhw7/tMoXGTVqFFWrVuX48eNXfNsiIr5m6D5CIv7RqFEjJk6cSGxsrL+r8qdlZmZy9913U79+fYoXL86zzz7r7yqJiPwjygiJXEOqVq3KU089RZMmTdi9ezcffPAB7du358EHH6Rhw4bMnz8fgCVLlvDvf/8bgM6dOzN27FgeeeQRGjVqxMCBA/F4PBw5coTq1asD8MYbb/DCCy/w+OOP07RpUzp16kR8fDwAu3btom3btjzwwAP06dOHNm3asH379lzrt3r1asqUKcNjjz3GwoULSU9Pz37s0KFDdO7cmRYtWvDAAw/w4Ycf/uH6Ro0asXv37uzn/7585MgR7r77brp3706TJk1ISEhgypQptGvXjgceeID77ruPjz/+GACXy8XIkSNp0qQJzZs3Z9CgQWRlZdGkSRM+//zz7G0PHjyYWbNmXZH3SETyFnWERK4hTqeThg0bsnbtWsqXL8+iRYuYNm0ay5YtY/z48YwePTrX5/3yyy/MmTOHFStWsG3bNr744ouLyuzYsYOJEyeyZs0aoqKieP/993G5XPTt25ennnqKlStX0rlzZ/bt23fJ+i1YsIBWrVoRGxtLdHQ0S5cuzX7s6aefpmnTpqxevZpp06Yxbtw4zp49e8n1f+T48eP07t2btWvX4nQ62bJlC3PnzmXlypUMGDCA//u//wNg/vz57Nmzh+XLl7Nq1SpSU1P58MMP6dixI4sWLQLg7NmzrF+/njZt2lz29ReR4OPwdwVExKpWrVoA5MuXjylTprBx40YOHz7M/v37SUtLy/U5DRs2xGazkT9/fsqWLUtycjKlSpWylKlduzb58+cH4JZbbiE5OZmDBw8CcPfddwNQp04dKlSokOs+9uzZw759+5g2bRoADz74ILNnz6Zjx44kJyezf/9+2rdvD0Dx4sX55JNPOH36dK7rL8fhcHDbbbcBULJkSV577TVWrlzJzz//zLfffktqaioAW7ZsoXXr1oSHhwMwYcIEAFJSUpg8eTJJSUmsWbOGe+65h6ioqMvuV0SCjzJCIteYyMhIwJsVefDBB/ntt9+oWbMm/fv3v+Rzfu8IABiGQW5T/3IrY7fbLyprt9tz3cf8+fNxOBw89NBDNGrUiDlz5nD48GE2bdqEw+HI3u7vfvrpp+xtXbj+94nhOfedlZWV/f/Q0NDsbe7Zs4cOHTpw9uxZ6tevT48ePbLL/V7md4mJiSQkJBAVFUXTpk1ZsWIFixcvpmPHjrnGJCKijpDINeq7776jUKFC9O7dmwYNGvDpp58C4Ha7r9g+brzxRkJDQ9m0aRPgnS908OBBS8cFvBmW1atXM2XKFDZs2MCGDRvYtGkTrVq14t133yV//vxUqVKFZcuWAXDs2DE6duxIRkZGruvPnDlDoUKF+O677wD45ptvOHHiRK51/PLLL6latSrdunWjdu3arF+/Pvs1qFu3LqtWrSIrKwuPx8OwYcNYvXo1AI888gizZ8/GNE2qVat2xV4zEclbNDQmco2qX78+H3zwAU2bNiUiIoJq1apRqFAhfv755yu2D4fDwRtvvMGLL77IuHHjuOGGGyhSpIglewSwdOlSbrzxRurUqWNZ/+STT9KiRQsOHjzI2LFjeemll5gzZw6GYfDKK68QHR19yfXPPvssw4YN4/3336dKlSpUqVIl1zq2bNmSdevW0bx5c0JCQqhbty7JycmcPXuWDh068Ntvv9G2bVtM06R27dp07twZgMqVK1OwYMGrcgsBEck7dPm8SJB77bXXePzxxylSpAjHjh2jdevWfPLJJwE/p+aXX36hc+fOrFmzhoiICH9XR0SuUcoIiQS5kiVL8thjj+FwODBNkxEjRgR8J2jixIksXLiQQYMGqRMkIn9IGSEREREJGL8Pi0+ZMuWiq2P37dvH4MGDOXv2LLVq1eKll1666KKKC2mytIiIiASEb7/9lo4dO3L48OFcHx84cCBDhgxh7dq1mKbJwoULL7tNdYREREQkICxcuJAXX3yRmJiYix777bffyMjIyL4HWdu2bVmzZs1lt6k5QiIiIuI3KSkppKSkXLQ+KirqovmKr7zyyiW3k5CQQHR0dPZydHR09k8J/ZGA7AjdX+dlf1fBb2ypmf6ugl8ZQR5/RoXoyxfKo0KTMvxdBf/a95O/a+BXttIl/F0Fv/po30if7s9zvKLP9jVrUV8mTZp00fq4uDj69u37p7eT25TnC++JlpuA7AiJiIhI3tC1a9dcfwvwr169WrRoURITE7OXT5w4kesQ2oXUERIRERELDx6f7Su3IbC/o2TJkoSFhbFz505q1qzJsmXLuOuuuy77PE2WFhERkYDVs2dPdu/eDcCYMWMYOXIkzZo1Iz09nS5dulz2+coIiYiIiIXb9F1G6O90RDZs2JD9/7fffjv7/5UrV+aDDz74S9tSRkhERESCljpCIiIiErQ0NCYiIiIWHoLn17eUERIREZGgpYyQiIiIWPjy8nl/U0ZIREREgpYyQiIiImLhzuXnKvIqZYREREQkaCkjJCIiIha6akxEREQkCCgjJCIiIhZuZYRERERE8j5lhERERMRCc4REREREgoAyQiIiImKh+wiJiIiIBAFlhERERMQieH5pTBkhERERCWLqCImIiEjQ0tCYiIiIWOiGiiIiIiJBQBkhERERsXAHT0JIGSEREREJXsoIiYiIiIUunxcREREJAsoIiYiIiIUbw99V8BllhERERCRoKSMkIiIiFh5dNSYiIiKS9ykjdE7tehXo3rsRISF2Dv2QwLhXVpCWlmUpc2/TWNo9UhdMyMhw8ua4NXy//1j24/nyhzF2ymOMHbHCsj4Q1G5QiW797/fG//1xxg9dSlpqZq5lnxnxEIe/j2fxrM8BGDS2IyXKFM5+vFjJ69m94xDD+s31Sd2vhNsb3ky3gc0ICXVwaP8xJrywkLSzucf/9OsP8/PB4yx+Z+NFjw1+qysn41N4a9jSq13lK6pO7fL06HY3ISF2fjp0gtHjP7ro839fo1vo0P4OTNMkI9PFG29+wsHvjwNw150VeaRDXUJC7MTHpzBy9CpSzmT4I5R/rHb9CnTrcy8hoXYOfR/P+BErSEu1vhaNmsXS/tF6mEBmhpM3x3zE9/sC55iv3eRWur3U3vt53/Mr43tPJ+2C9+tSZWw2gz7juhB7ZyUAvly7i7cHvQfAHc1uY+DUniQcOZm9nWfuf5X0s9fuZ+H2uyvRbUATb5wHjjNh8OJLtn1Pv9qOn7+PZ/HMzdnrWnSsQ9N2tQgNC+GHPb8xYfBinE63r6p/1WiOUJApeF0kzw5uxcv/WcTjD7/JsaOneLzPvZYypcoUpkfcfQzqP58nu0xj/szNvDjqX9mP3173Jt6Y0YPSZYv4uvr/WMHrI3l6eFuGD5hPj1YTOHbkFN36N7moXOly0Yx6pzsN7q9qWf/KMwvo034SfdpPYuKwpZw9k86kV1b6qvr/WMFC+Xj6tYcZ0Xs2Pe97neO/JtHtuRYXlSt9Ywwj5/aiQYtbc91OuyfuoWqtcle7uldcwYIRPPdMc14cvoyuPd7h2PHTPNH9bkuZ0qUK0atHQ54btJCevd9l7vwtvDy0DQAVKxSjX+/GDB2+lO7/nsGvvyXxeLe7/BHKP1bwukieGdqa4c8vpEe7yRz/7TTd4+6zlClVtjA9+jVmUL959H5kKvOnb2Lo6w/7qcZ/XcEiBXhmSg+GP/IGPWq8wPFDJ+j+8r/+dJl7O9anVIVi9Ko9iCfrDCH2zko0aHM7ALfcUYEP/u8jetcbmv13LXeCCl6fj6dfaceIp+bRs/k4jh9JotszTS8qV7p8NCNn9qBB01jL+nqNq9Dqkbr8p/t0ej0wgdDwEB7seqevqi9XiM86Qj/++CNvvvkmQ4cOZdiwYbz55pvs3r3bV7v/QzXvKM+BfUc5+msSAKuW7KBRE+sH3ul0Mf7VVSSdPAvA9/uPcn3h/Dgc3pfwwX/VZvTLyzmZeMa3lb8CatSrwME9v3H0F++3uNXvb6dRLif7BzrW4eNlX7F53Xe5bsfhsPPMK+2Y+tqHJMYnX9U6X0k1GlTk4O5fOXo4EYBVc7fQsHX1i8q17Fyfjz/4gs2rv73osWp1bqTmXZVZPX/rVa/vlXZ7jXIcOHCc346eAmD5qq+5t1EVS5ksp4sxEz4iKSkVgAMHj1Po+nw4HDYa31uFD9fuIj4+BYBZcz7nvYXbfRvEFVKjzo0c2Pvb+bZg8Zc0uuDk58xyMWHEyuy24OA+a1twravRqCoHdv7E0R/jAVj1zgYa/avuny5js9sIjwwjJCyEkDAHIaEOsjKcANxS5yZuu/sWJm1+ibHr/kvV+pV8GNlfV6N+BQ5+d4SjP3vbvlULttGw5W0XlWvZqS4fL93J5jXWc9a9raqz5N3NnE1OxzRNJg1bxoYVX/uk7lebG8Nnf/7mk6GxefPmsXDhQpo0aUJsrLdROXHiBEOGDKFVq1Z0797dF9W4pOiYgpzIceI+kZBCvvzhREaGZg8PxB9LJv7Y+TL/fup+tm0+gMvlve3UoAHzfVvpKyi6WEFOHM8Rf3wK+QqEE5kvzJIifvNVb5bntjo35rqdJm1rknQihS0b9l7dCl9hRYpfx4ljp7OXE48nk69ABJH5wyzDY78Pd91Wr4Ll+YVioug1tDWDur5N807WE0ogiI4uQEJiSvbyiRNnyJ8vzPr5j0/J7ugA9P53I7Zs+wGXy0Opktfz06ETjBjWlmJFC/LToRNMnrre53FcCdFFo0jMEWd2W5AvNHt47KK2YEATtm063xZc66JLFSLxt6Ts5RO/JZGvYCSRBcKzh8f+qMzHczdzV5vbmXdwAnaHja82fMf2j74BICXpLOsXbGHLyp1UqVuBYe/158m6g0k818m+1hQpVpATOd7LxEu0fW+NWAFc3PaVuqEIB3fnZ/i0bhSOKcB3Ow8zfcxHvqm8XDE++Qoze/Zs3nvvPXr37k379u1p3749vXv3ZsGCBSxcuNAXVfhDhi33Hqknl2nz4eEhDH6lHSVKFWLcq4Ez/PNHDCP3+N2ev9awt+lcnwVTP7sCNfIt2yXef/ef+LEdu8PGC//3KFOHr+DUicDLBsKl4/fkEn94WAgvDmpNyRLXM3q8t8F3OOzUq3MTYyeupWfvmSSdSuXZ/hcPLwQC26WOhVxei7DwEAaN9LYF48+dKAPBpT/vnj9V5tH/PsjpxDN0KN+XRyoNoMD1+Xmor/f9Ht7pDbas3AnAnq3fs3f799RoVDXXbV0LLhnnn2z77CF2qte7iZED5tOv/WQKFIzksVymFQQij2n47M/ffNIRcjgcuFyui9ZnZGQQEhLiiyr8oRPxyRQqUiB7uUh0FCnJ6WScS/f+LrpoFOPf7obH42Fgn9mkXmIybaA5cfw0haJzxB8TxZnkNDLTnX/wLKsbKxfH7rCxa8ehq1HFqyrhtwviL1qQM6fTyEzP+oNneVWILU2xUoXoOegBJq0aQPNOdbi7xa08NbL91azyFRWfkELhQvmzl6OLFCDlTDoZmdb3Pya6AJMmPIrHYzLguQWknvvGfPLkGb7ccYhTp1IxTVizbje33FzSpzFcKQnxyRQqcv61KBIdxZnkdDJzaQsmTO+Ox2Py3JOzAqotSPg1iUJFr8teLlLies4knSUzx+T4PypTv1Ut1s3ZhMvpJi0lnY/nfc6td91MvoKRdHi2pWVfhmHguoYnDiccu/DYjzp37P+5ti8pIYUtn+wlLTUTl9PNhpVfU/nWMlerunKV+KQj1KtXLx588EEGDx7MxIkTmThxIoMHD6Z9+/b06tXLF1X4Qzu3/8jNVUtSonQhAFq2qcnWzQcsZQpEhTP2ra7877P9vDpkCVmZF3fsAtXOLT9QuVrp7Cu/WvyrNls/3feXthFbqxzfbv/palTvqvvq84NUrl6WEjd4J7o3f6QOWz/Z86eeu//rn+ly5wjiWo4nruV4Ppy/jY2rv2XifxZdzSpfUTt2HubmyiUoWeJ6AB5ocRv/2/qDpUyBAuFMGNOJTZ8fZPjIFWRlnf/8b9x8gDtqlyeqQDgADepX5MDBwLmCKqed236kctVS2W1Bi4dqsXXTfkuZAlHhjJn6GJ9/up+RgxYHXFuwc8NuKte+kRI3FgWgxeON2Lr66z9d5odvfuautncAYHfYqdOiOvu+/JH0M+k88MR93Nm6FgA3VitDpVrl2fHJLl+F9pd99b/vqXxraUqU9bZ9zR++g61/YWj/87Xf0aBJVULDvLNM6t57Cwe/O3JV6ipXj0/mCD3wwAPUrl2brVu3kpCQgGma1KpVi759+1K0aFFfVOEPnT6VxpjhKxjyajtCQuwcPXKK0S8vo0Ll4jz93wd4sss0WratRXTRgtS/uzL1766c/dzn4uZwJiXdj7X/55KTUhk3ZDGDx3XEEWLn2K9JjP7vB1S4pST9X2pDn/aTLruNkmULE3+NzgO4nOSTZxn/3PsMmtzFG/8vJxnzzAIqxJbiqZHtiWs53t9VvKpOJ6fx+tgPeWnIgzgcdo4eO8XI0aupWKEYAwc0pWfvd2nVsjox0VE0qF+BBvXPz5F65vn32Lr9R6KjCzBhTCcMwyA+IYXR4wJznkTyqTTGvrycIaPaez8LR04xethSKtxcnAGDW9H7kam0fOh2oosVpH7DytRveL4teL73bM4kX/ttQfKJM4zt9Q5D5sbhCHVw7KcERj8xjQrVb2DA5O70rjf0kmUAprwwjz5jOvPOVyPxuE2+/mwPC8etxuMxGfbwBHqP6UznQW1wu9y82nUyKecmlV+LkpNSGT9oMYMmPJLd9o15YSEVqpTkqeFtiWv7xh8+f9WCbeQvGMkbH8Rhs9v4Ye9R3nktsG6dcSnXwiRmXzFM0wy4+0feX+dlf1fBb2yXuL9FsDCCPP6MCtH+roLfhCZdu5dh+8S+wMy4Xim20iX8XQW/+mjfSJ/u76tffDfEV6PMLz7bV250Q0URERGxcAfRbQaDJ1IRERGRCygjJCIiIhbXwmXtvqKMkIiIiAQtZYRERETEIpiuGlNGSERERIKWMkIiIiJi4TaDJ08SPJGKiIiIXEAZIREREbHwBFGeJHgiFREREbmAMkIiIiJioavGRERERIKAMkIiIiJioavGRERERIKAOkIiIiIStDQ0JiIiIhYeTZYWERERyfuUERIRERELdxDlSYInUhEREZELKCMkIiIiFrp8XkRERCQIKCMkIiIiFvrRVREREZEgoIyQiIiIWLhN3UdIREREJM9TRkhEREQsdB8hERERkSCgjJCIiIhYeHQfIREREZG8TxkhERERsdAcIREREZEgoI6QiIiIBC0NjYmIiIiFbqgoIiIiEgQCMiM05L3Z/q6C36SZof6ugl9lmQH5kb1iMswQf1fBb8o4kvxdBb9K8YT7uwp+5SF4MhS5G+nTvelHV0VERESCQHB/vRYREZGLuHVDRREREZG8TxkhERERsQimOVnKCImIiEjQUkZIRERELDRHSERERCQIKCMkIiIiFvrRVREREZEgoIyQiIiIWHj0W2MiIiIieZ8yQiIiImKhOUIiIiIiQUAdIREREQlaGhoTERERC49uqCgiIiKS9ykjJCIiIhZu/eiqiIiISN6njJCIiIhYaI6QiIiISBBQRkhEREQsNEdIREREJAgoIyQiIiIWmiMkIiIiEgSUERIRERELtzJCIiIiInmfMkIiIiJi4dFVYyIiIiJ5nzJCIiIiYqE5QiIiIiJBQBkhERERsfCYmiMkIiIikuepIyQiIiJBS0NjIiIiYuEOojyJOkLn7NoOi2facDmhVDl4bICHiHzWMuuXG2xYYRAaCsXLmHTqY5I/6vzjSQnwan8bL77loUBB39b/Stuz3WTlTA8uJ5QoBx0H2IjIZx0z3rjcw+YVJiFhULS0Qfs4g3wFAnNcee8Xbj6a6cblhOLlDP7V30H4BfHu/p+bdXPdGDaIyA/tnwqhSAmDtDMmiye5OPqjSWg43N7Yzp2t7X6K5O858IWLde9m4XaaFC1no03/cMIjrfHv3eJi/dysc/EbPPhUGIWL2/C4TVa9lcmh3W4AKt7uoOnjoRhGYHwWvtlusGiGHafToHQ5kx5Puy469tcts/HJCjuhoSYlyph0iXOTPwqyMmHWJDuHDhiYJpSvbNI1zk1omH9iuRK+226yYqaJywkly0GnAcZFx/5ny002nTv2i5WGfwXwsZ9TsLV7gWjlypW89dZbOJ1OHnvsMR555BHL43v27GHo0KE4nU6KFy/O6NGjiYqKusTWvIKny/cHzpyGmWNt9B7i4ZXpHqKLmSyeYf1g7/8G1iw0eGaUhxff8hB7O8yZeP7l2/KxwWvP2jh9MvAPiDOnTeaN89B9iI3B0+0ULm6wcqZpKXPwW5NPFpn0GWXj+Tft3HI7vDfR46ca/zNnT5u8P85Fl8EOnn8nlELFDFbPdFnKODNN5o920XVICE9PDqVKHRvLpnjLLJ/qIiwcBk4Noe/4EPbv8LB3u9sfofwtqckmS8Zn0nFQOP3fzkehYjbWzcy0lHFmmiwanUGnweHETYqk8h12Vk/xlvlmg4sTR0z6vhlJ3ORIDu92s+fzwIg/5TS8PcZB36EuXp/hJKa4yfvTrZ3Yvd8YrF5o54XXnIyY4uLW2iYzJ3i/Q66Yb8fjhhFTXLwyxYUzE1a+F1id4JzOnDaZO86kxxCDodNtFC4OKy5x7PcdZfCfN21Uud1gwUTzElsMHMHW7l2OxzR89vdnxcfHM378eObPn8/y5ct5//33+eGHHyxlXnnlFfr168eKFSsoV64c06dPv+x21REC9nxlcEMlKFrSu3xPS5PtG7zf8H738/cGN1c3KRTtXa5xp8m328HlhNMn4eutBk8NzxsHxP6vTMpUhJiS3g/onS0MdmwwMXO8IL9+b1LpNoPro71lbr3T4Lvt4HIGXoN48CsPpSvaiC7pPRzqtbTz9aceS7yec29tRqp3XWY6hIR61x35waTGvXZsdgNHiMHNtW3s+jxwPgvff+WiZEUbRc7FX7tFCN9+6vrD+LPSwRFy/jFnhjeD4HKC2wWOUJ+G8Ld9t9NG+Uomxc4d+41autm6wWY59g9/b1Cluif72K9V38PX2w1cTqgU66FVJzc2G9jsUPYmk8R438dxpez/CsrmOPYbtDD4cgOWz8Iv30Ol28hx7BOwx35OwdbuXUtSUlI4cuTIRX8pKSmWclu2bKFOnTpcd911REZG0qRJE9asWWMp4/F4SE1NBSA9PZ3w8PDL7l9DY0DSCShU5PwH+fpoSE8zyEgjO0VerrLJ+uU2TsabFC4K/1tr4HIanE2B6wpDn6GBc+K7nNMnzjdyANdFQ0YaltejbCWDTcs9JMWbFCpqsH2didsJqSlQsLCfKv43nU40uS76/HLBIt5YM9Mg/Fy8YREGD8U5eONpJ/mivCf/uLHes33ZSgZfrXdT7hbvyXHX/9zYAygpkHzCpGCR8+93VBGDzDRvZy880rsuLMKgVVwY055JJzLKwOOBJ8ZEAFDjPgd7PnfxepdUPG64qbqdyncERtNy8gQUij5/7BfK5dgvX8lk3TI7ifFuihSFTets2cd+bK3zz02Mh7VL7HTr77pwNwHj1Aksx0Jux/4NlWDjcrKP/W3rvB3gQDz2cwq2du9yPD7Mk8yaNYtJkyZdtD4uLo6+fftmLyckJBAdff4DGhMTw65duyzPeeGFF+jWrRuvvvoqERERLFy48LL790lrdfTo0T98vESJEr6oxiWZl+jD2HKczCrGwgOPmkx+2YZhwJ1NTPIVMLO/Fecl5iW+3OR8PW6KNWj6iME7L3swbFDnfoPIAgTk63Gp99/IEe+xQx4+nu9i4NRQipQw2LzcxawRTp6eHMIDPR2sfMfFuDgnUYWgYnUbh/cGzjfES77fOdrB44fcfDo/i35TIylc3MbW5VkseCWDPpMi2DA/i8gogxfm5cOVBfOGZ/D5kizubHvtp4X+TOyVq5m0edTNxJccGAbc1cRDvgIm9hyf9UMHDSa+5OC+1m6q1wmc9/5Cf/bYb/YITHvZxLCZ1D137NsD8NjPKdjavWtJ165dadOmzUXrL5zbY+byJuWci5iRkcGgQYOYNWsW1apVY+bMmTz//PNMmzbtD/fvk47Qv//9bw4fPkxMTMxFgRiGwfr1631RjUsqFAOH9huAt26nEyEyv0lYjoxaRhpUjDVp0NRbJvkULJtlkK+AHyp8lV0fDYf3n3+fkhMhMj+Ehef4wKWZ3FTNoG5T7xkj5ZTJ6tkmkQH4elwXY/DLAWu8ERfEe2CnhxtusVGkhHdd/ZZ2Vkxzk5binTDb8nEHkecmTG5Y6MouFwiuizY4kiP+lESTiPwQmiP+H75yU+YWO4WLe9/vO1qG8OHbWaSlwN4tblr2CsURYuAIger3ejNEd7b1eSh/WeFo+HH/+ThPJUK+AiZhEefLpKdB5Woe7m7m7TEnn4LFs+zkP/dZ3/apjVmT7HTu46Zeo8DODHuP/fPLlzr2K1SDejmO/VWzCfi2MNjavctx+/CGilFRUZed0AxQtGhRduzYkb2ckJBATExM9vLBgwcJCwujWrVqADz88MNMnDjxstv1Se5rwYIFlCtXjtdff50NGzZY/vzdCQKoUtPkx/0Q/5t3+bPVBrfVtXbYTp8H+1GIAAAgAElEQVSEMc/ZSPcOPbJqnkHte0wC5MKYv6RyTYOf90PCb97X4PPVJrF1rYEmn4T/e85D+rk5I2vnm9S8xwiYK4VyqljDxs/7PZz4zXsS2/ahmyp1rYdGyZts/LTbw5lT3ni/2+qhUFHIV9Bg64du1szxDoecOWWyfY2b6vcEzvS7m2rY+XW/h8Rz8X/5oZPKdazfkYrfaOPwbjdnT3nL7Nvq5vqiBvkKGpS40cZ3m73xu10m+7e7KFU5MMYGY2t6+HGfwfFzx/6GVXZq1LV2Zk6fhFcHhmQf+8vn2anb0INhwBebDOa8aWfgSFfAd4IAbq7p7Qj9fuxvXm0SW9daJvkkTHzOzD7218w3qXUPAXns5xRs7V4gqlevHlu3biUpKYn09HTWrVvHXXfdlf142bJlOX78OD/99BMA69evJzY29rLbNczcck1Xwa5du1i0aBHDhw//x9vafPimK1Ajq11fwJIZNlwuiCkO3Qd6SDwOs8Z7L4cH2LDcYMNK7yTqClW8l89feJlsjyZ2xi90X7XL59NM3ww37PnCexmp2wVFisOjA22cPAYLJnh4/k3vSW7TCg+bV5qYHihfxaBdH4PQsKvbIGSZVyeJue8LNx++68btgsLFDTo+6+DkMZNFE108Pdn7mv9vpZv/rXRjd0BkAWjT20GxsjYy0kwWjHGReNQEExo9bKdmo6vTEcgwr04O/sCXLj5+Nwu3y6RQMRsPPRvOqWMelv5fJnGTvBOFtq3MYvsqJ3aHQUQBg5ZPhlK0rJ20FO/l80d/dGOzGZS/zU6zHqHYHVf2s1DGkXRFt/e7b78wWDjDjstpEFPC5N8DXSQcN5gxzs6Ic1cGfrzce/m8aULFKh66nLtEfuBjIaSlwvWFzzejFaqYdO175a+aS/FcftLnlbDni3OXz5879rsMNEg8BvMnmPznTW8Hf+MKk03njv0bq0B7Hxz7Hq5+Z+NabfcAmpTbe9X3kdNTX3f02b4mVl/wp8uuXLmSqVOn4nQ6adeuHT179qRnz57069eP2NhYNm7cyNixYzFNk8KFCzN8+HBKly79h9v0WUfoSroaHaFA4auO0LXqanWEAsXV6ggFgqvVEQoUvuoIXat80RG6lqkjdPUE91lFRERELuIxA2d4/58KnkhFRERELqCMkIiIiFi4g2goUhkhERERCVrKCImIiIjFX/kNsECnjJCIiIgELXWEREREJGhpaExEREQsdPm8iIiISBBQRkhEREQsgulO3soIiYiISNBSRkhEREQs3Lp8XkRERCTvU0ZIRERELHTVmIiIiEgQUEZIRERELPQTGyIiIiJBQBkhERERsdB9hERERESCgDJCIiIiYqE5QiIiIiJBQBkhERERsdB9hERERESCgDpCIiIiErQ0NCYiIiIWmiwtIiIiEgSUERIREREL3VBRREREJAgoIyQiIiIWmiMkIiIiEgSUERIRERELZYREREREgoAyQiIiImKhjJCIiIhIEAjIjFDdcH/XwJ+y/F0Bv7Lh8ncV/CzD3xXwoxB/V8CvnGamv6vgV6c8wfzZ9z1lhERERESCQEBmhEREROTq0Z2lRURERIKAMkIiIiJioTlCIiIiIkFAHSEREREJWhoaExEREQsNjYmIiIgEAWWERERExEIZIREREZEgoIyQiIiIWCgjJCIiIhIElBESERERC1MZIREREZG8TxkhERERsdCProqIiIgEAWWERERExEJXjYmIiIgEAWWERERExEJXjYmIiIgEAWWERERExEJzhERERESCgDpCIiIiErQ0NCYiIiIWmiwtIiIiEgSUERIRERELTZYWERERCQLKCImIiIiFafq7Br6jjJCIiIgELWWERERExMKD5giJiIiI5HnKCImIiIiF7iMkIiIiEgSUEfqLTBMGjTKoUA66dQiiafUEd+zgjf+/o6BCOejewd+18T3Fnzfj37TVYOLbdrKcULG8yUvPucmfz1pm/hIbC5baCA+FcmVNBvV3UzAKklNgxHg7+38wiAiHB5t56NTW459A/qat2+xMeycMZ5ZB+fJunh+YQb4L4l+8JISly0IJCzMpU8bDgKcyiIryPrZ0eQirV4eQmQWVKnp47tkMQkN9H8eVpvsISa5+PAzdB9hY+2nwfEB+F8yxgzf+bgNgzaf+rol/KP68GX/SaRjymp1xL7tYOcdFqRImE6ZZTwtffG0wY76Nt8e6WDTdRYM6Hl4aYwfg9cl2IiNg2bsu5r3p4vPtBhu3BE4bcfq0wajXwxk+LJ25s1MpUcLD1LfDLGW++trOgvdCGTc2jelvp1HnDhdjxoUDsGmTgyVLQxg3Jo1ZM9LIzIRFH+SBXlCQ8VlH6JNPPmHOnDn88ssvlvXvv/++r6rwjy1YZtCmmUmThsGXDQnm2AHmL4M2zaBpQ3/XxD8Uf96Mf+uXBlUrm5Qt5V3+VysPH35is9xDZu8Bgzo1TYrFeJfvbWCycauB0+l9rGVjD3Y7hITAXXVMPt4YON+vv9xhp3IlD6VKeQNu3crJJ+tDLPEfPGijZk03MdHelXc1cLFlqwOnE9Z+7ODh9k6iosBmg2cGZHJ/Y6c/QrniTNN3f/7mk0/smDFjmDt3LocPH6ZDhw4sX748+7H33nvPF1W4Igb3N2nV5Bp41/wgmGMHGNIfWjfxdy38R/HnzfiPJxgUiz5/XBeNhrOpBqlp58tUvdnki68Njh73Li//yIbTaXA6BardYrLqYxtOF6SlwcebDBKTfBzEP5CQYCMm5vxQXnS0SWqqQVqO+G+u7OGrr+0cP+7NdH20JgSn0yAlxeDXIzZOnTYY+HwE3XpEMnNWKPnzB287Gah8Mkdo48aNLF26FIfDQefOnenevTuhoaE0a9YM81roDoqIBCHPJZpfW46vyLVuNenV1U3/IQ5shsmDzU0KRpmEOOCZJ92MfcvOv3o4iC5sUreWyTffBc7Q2J+J/9Zb3TzWJYvBQyOw2aBZMydRUSYOh4nLZbBjp51Xh6cTGgqvjgrnnelh9I3L9E0AV1EwXTXmk46QaZoYhvdFveGGG5g6dSrdunWjUKFC2etFRMS3iseY7N53/qyfkAhRBUwiI86XSU3zdobatnABcDIJJs+wUTAKjifA0728E6cBZsy3UaZk4Hy5LRrjYd++86fBxBMGBQqYROSIPy0Nbr3VRYvm3iGvpCSDGTPDiIqCIoU9NLjTlT25+v7GTmbNts4xkmufT4bGmjZtSufOndm1axcAFSpUYOLEifTv3/+iOUMiIuIbdW832bXX4Ocj3uVFK2w0rG/tyCQkQvf+Ds6mepenzrbRrJEHw4CFK2xMnuE9jZxMgsWrbDS/L3A6QrfXcrN3n50jR7xfyFesDKF+PZelTGKiQf8BkaSei3/2nFDubejEMODuu1x8ttFBZqZ3rsvmzx1UruT2dRjyD/kkIxQXF0fNmjXJl+OaxJo1a7JkyRJmzJjhiyqIiMgFCl8Pw59388yL3sm/pUuYvPJfN3v2GwwbbWfRdBflysDjnTw88qQDjwk1Yk3+85T3ZN/jEQ//fcVOm8e8p5InH3NTtXLgdISuv97khYEZDB0WgdMFJUuY/PeFdPYfsDF6TDjT306jTBmTTh2z6NUnEtNjEBvron8/79DXg62dnDlj0LNXJB63QYUKbvo8meHnqK6MYBoaM8wAnKTjOn6Tv6sgfmLTHR8kSDlN1+UL5WGnPHmjg/F3FSt51Kf7i13xos/2tbvVSz7bV250Q0URERGx0A0VRURERIKAMkIiIiJiEXiTZv4+ZYREREQkaCkjJCIiIhbBdNWYMkIiIiIStJQREhEREQtlhERERESCgDJCIiIiYhFEF40pIyQiIiLBSxkhERERsdAcIREREZEgoIyQiIiIWAXRJCFlhERERCRoqSMkIiIiQUtDYyIiImKhydIiIiIiQUAZIREREbEwNVlaREREJO9TRkhEREQsNEdIREREJAgoIyQiIiJWygiJiIiI5H3KCImIiIiFrhoTERERCQLKCImIiIiVMkIiIiIieZ8yQiIiImKh+wiJiIiIBAF1hERERMTK9OHfX7By5UqaN29O48aNmTdv3kWP//TTT3Tu3JlWrVrx+OOPk5ycfNltqiMkIiIi17z4+HjGjx/P/PnzWb58Oe+//z4//PBD9uOmafLkk0/Ss2dPVqxYwc0338y0adMuu13NERIRERG/SUlJISUl5aL1UVFRREVFZS9v2bKFOnXqcN111wHQpEkT1qxZQ1xcHAB79uwhMjKSu+66C4BevXrlut0LqSMkIiIiFr6cLD1r1iwmTZp00fq4uDj69u2bvZyQkEB0dHT2ckxMDLt27cpe/uWXXyhSpAjPP/88e/fupWLFigwZMuSy+1dHSERERPyma9eutGnT5qL1ObNB4B36upBhnO+wuVwuvvjiC+bOnUtsbCwTJkxg1KhRjBo16g/3H5AdIZumNolIkLEbwd3uOYPpDn/XAh++3BcOgV1K0aJF2bFjR/ZyQkICMTEx2cvR0dGULVuW2NhYAFq2bEm/fv0uu93gPrJEREQkINSrV4+tW7eSlJREeno669aty54PBFC9enWSkpLYv38/ABs2bKBKlSqX3W5AZoRERETkarr2bqhYtGhRBgwYQJcuXXA6nbRr145q1arRs2dP+vXrR2xsLJMnT2bw4MGkp6dTrFgxXn/99ctu1zBzG3S7xnmOV/R3FUREfMqDx99V8Ktj7lR/V8GvSpc85tP93TD7NZ/t63CX5322r9woIyQiIiJWAZci+fs0R0hERESCljJCIiIiYqWMkIiIiEjep4yQiIiIWPnwztL+poyQiIiIBC1lhERERMQi8G6s8/cpIyQiIiJBSxkhERERsVJGSERERCTvU0dIREREgpaGxkRERMRKl8+LiIiI5H3KCImIiIiFocnSIiIiInmfMkIiIiJipYyQiIiISN6njJCIiIhY6aoxERERkbxPGSERERGx0hwhERERkbxPGSERERGxUkZIREREJO9TRkhERESslBESERERyfuUERIREREr3UdIREREJO9TR0hERESClobGRERExMLQZGkRERGRvE8ZIREREbEKooyQOkJ/kWnCf0dBhXLQvYO/a+NbwRw7KH7Fnzfj37gVJkyzkeWEiuVh+PMe8uezlpm32GD+UoOwMChfxmTwAJProuB0CgwfZ7D/B4OIcGjTzOSRhwLrDLptm4Pp74ThzILy5T08MzCdfBfEv3RJCMuXhRIWBmXKeOj7VDpRUd7Hli8P4aPVoWRlQYWKbp55NoPQUN/HIX+fhsb+gh8PQ7cBsOZTf9fE94I5dlD8ij9vxp90GgaPsjFhuIfVcz2UKmEybqr1suntX8H0BQbTx3lYMt3DXXVg2BjvqeO1SQaREbBilof5b3nYvN3gsy3+iOTvOX3aYMzr4bw4LJ13Z6dSvISHd94Ot5T55ms7778XxuixaUx9O5Xad7gYPy4CgM2bHCxfGsrrY1J5Z0YqmZkGiz9QLyjQ+KwjdPjwYeLj4wFYtGgRI0aM4MMPP/TV7q+I+cugTTNo2tDfNfG9YI4dFL/iz5vxb/nSoGplKFvKu9yhtcnqTwzMHEmdvQcN6tQ0KRbjXb7vLpPPtkCW0/vYA/eb2O0QGgJ31TVZtzFw7j+zc4edipXclCrlAeCBVlmsXx9iif/gQTs1arqIjvauvLOBk21bHTid8PHHIbRrn0VUFNhs0H9ABo0bO/0RivwDlx0amzFjBpMmTcLtdlOiRAkqVaqU/VexYkVKlSp12Z28++67zJkzB4/HQ506dTh27BiNGzdm8eLFHDp0iD59+lyRYK62If29/277yr/18Idgjh0Uv+L3/pvX4j+WAMVizp/1i0bD2VSD1DSyh8dibzaZu9jG0eMmJYrB0o8MnE6D5BSodrPJynUG1WNNsrLg440GDgcEygSThAQbMTnij442SUs1SEsje3iscmU3S5eGEn/coGgxk7VrQnA6DVJSDI4csXH6tMELz0dy8qRBbKybnk9k+CmaKyuYrhq7bEdo6tSpvP7661SrVo1ff/2VgwcPcuDAATZt2sT3338PQIUKFViwYMElt7F48WI+/PBDEhMTadmyJdu2bSMsLIz27dvTrl27gOkIiYjkJaYn9/W2HGMFtW6F3o+Z9Btsw2bzzgMqGGUS4oCBvU3GvGXQroeN6EJQr5bJ198FTkbIvMTJPmf81W5106VLJi8OjcRmgybNsigQ5cHhALcLdu508PLwNEJD4fVREcycHkbvuEzfBCBXxGU7Qvnz5+eee+7B4XAQExNDzZo1LY8fOXIku0N0KR6Ph9DQUEqWLEn37t0JCwvLfsztdv/NqouIyD9RvCjs2mfwewYnIRGiCphERpwvk5oGtW41eaiFt0xiErwxw6BglDej9HQvk+uivI+9M9+gTKnASSXExJjs23e+45Z4wqBAAZOIHPGnpUG1W100a+4d8jqVZPDuzDCiokwKFza5805Xdvbo3sZO5s4OA/JAR0g/sXHeE088waJFiy75eKlSpWjY8I8Hzu+//34effRR3G43ffv2BWD//v106tSJZs2a/cUqi4jIlVDvdpNde+HnI97l91cYNKpv7cgkJEK3/jbOpnqXp8w2aH6viWHAwuUGk2Z4T5iJSfDBKoMW9wZOR6hmLRf79tk5csR7Kly5MpR69axzfE4m2nhmQD5Sz8U/d04YjRq6MAxocJeTjRsdZGZ6s0v/+9xBpUr6ch9oLpsRGjVqFE6nk82bN9OgQQNuvvlmKlWqRETOLvNlPPXUU3z55ZfY7fbsdaGhofTt25e7777779VcRET+kcLXw4gXPPQfasPlhNIl4dX/evhuPwwdbWPJdA/lysDjnUw69rLhMaFGrMmg/t7OTs9HTV54xUbrx7wTrPs8ZhJ7s5+D+guuv95k4MAMXh4WgcsFxUt4eP6FdA4csDFuTART306ldBkPHTpm0rdPPjweqBrrpm8/7zygVq2dnDlj8GSvfHjcUKGCh15Ppvs5qiskcPqz/5hhmpcaJfX69ddf2b9/PwcOHODAgQPs37+fo0ePUqpUKdauXeurelp4jlf0y35FRPzFwyUm9ASJY+5Uf1fBr0qXPObT/ZWfMM5n+/qp/9M+21duLpsRKl26NKVLl6Zx48bZ69LS0jh48OBVrZiIiIj4SRBlhP7WfYQiIyO57bbbrnRdRERERHxKP7EhIiIiFsF0HyH9xIaIiIgELWWERERExEoZIREREZG8Tx0hERERCVoaGhMRERErDY2JiIiI5H3KCImIiIiFLp8XERERCQLKCImIiIiVafi7Bj6jjJCIiIgELWWERERExEpzhERERETyPmWERERExEJXjYmIiIgEAWWERERExEoZIREREZG8TxkhERERsdAcIREREZEgoIyQiIiIWCkjJCIiIpL3qSMkIiIiQUtDYyIiImKloTERERGRvE8ZIREREbHQ5fMiIiIiQUAdIREREQla6giJiIhI0NIcIREREbHSHCERERGRvE8ZIREREbHQVWMiIiIiQSAgM0K/uM74uwp+E+8O93cV/CrS5vR3Ffxqf1ZRf1fBb0INt7+r4FexofH+roJfTUuq7+8q+NWokj7eoTJCIiIiInlfQGaERERE5CpSRkhEREQk71NGSERERCx01ZiIiIhIEFBHSERERIKWhsZERETESkNjIiIiInmfMkIiIiJiocnSIiIiIkFAGSERERGxUkZIREREJO9TRkhERESslBESERERyfuUERIRERELXTUmIiIiEgSUERIRERErZYRERERE8j5lhERERMRKGSERERGRvE8ZIREREbHQVWMiIiIiQUAdIREREQlaGhoTERERKw2NiYiIiOR9ygiJiIiIhSZLi4iIiAQBZYRERETEShkhERERkbxPGSERERGxUkZIREREJO9TR0hEREQsDB/+/RUrV66kefPmNG7cmHnz5l2y3GeffUajRo3+1DY1NCYiIiLXvPj4eMaPH8+SJUsIDQ2lQ4cO3HHHHdx0002WcomJibz22mt/ervKCImIiIiV6cO/P2nLli3UqVOH6667jsjISJo0acKaNWsuKjd48GDi4uL+9HaVETpn+zYHM98Jw+mEcuU9DHg2nXz5rGWWLw1hxbJQQsOgTBkPffqlExXlfexfbfNTuMj5d7T9vzJpdJ/LhxH8M99sN1g0w47TaVC6nEmPp11EXBD/umU2PllhJzTUpEQZky5xbvJHQVYmzJpk59ABA9OE8pVNusa5CQ3zTyx/x85tduZND8XlhDLlPfR+JpPIC+L/cKmDNctDCA0zKVnGpEffTApEQepZeGtsGL/9asP0wN33u2jTwemfQP6mA1+4WPduFm6nSdFyNtr0Dyc80pq03rvFxfq5WRg2iMhv8OBTYRQubsPjNln1ViaHdrsBqHi7g6aPh2IYfzXp7T/7vnDz0UwnLicUL2fQvn8o4fms9f/uf27WzXVmx9/+qRAKl7CRdsZkySQnR3/0EBoOtzd2UL914DStX1zQ9vW/RNu3clkoYWFQ+lzbV+Bc2/dw2/wUydH2PRRgbd/xnSnsmR+Px+khqmw4NZ4sRUik3VLm6PZk9i1MwDAgJL+d6r1Kkr/Y+QYuLTGLjf/9kUZjKhAWFTjv/bUiJSWFlJSUi9ZHRUUR9ftJFkhISCA6Ojp7OSYmhl27dlmeM3v2bG655RZuvfXWP71/ZYSA06cNxo4OZ8iwdKbPSqVYcQ8z3gm3lPnmazsL3wtj1Jg03pqWyu13uJg4LgKAX3+1kT+/yVvTUrP/AqkhSDkNb49x0Heoi9dnOIkpbvL+dGtDsPcbg9UL7bzwmpMRU1zcWttk5gTvAb9ivh2PG0ZMcfHKFBfOTFj5nj23XV2Tkk/D5DFhDHwxg/97N52ixT3MeyfUUua7b2wsez+EF0dnMGZqBjVqu5k63tsQvvduKIWjTca/k86oyemsW+ngwN7AObRSk02WjM+k46Bw+r+dj0LFbKybmWkp48w0WTQ6g06Dw4mbFEnlO+ysnuIt880GFyeOmPR9M5K4yZEc3u1mz+duf4Tyt5w9bbJwXBadB4fy3DvhFC5m46OZ1o6sM9NkwegsugwJZcDkcG6pY2f5FG+ZlVOdhIXDs1PDiBsfxv4dbvZuD4z4T582GDc6nMHD0nnnXNs384K279uv7Sx6L4yRY9KYfEHbd+RXGwXym0yelpr9F0htX2ayi51vHuGOZ8vQ+P8qka9oKHvmHbeUcWd62PHGr9zxbBkajalA8VpR7JpxLPvxXzaeYvPQn8g4FThx/xmG6bu/WbNmce+99170N2vWLEudTPPi9FHOL1wHDx5k3bp19O7d+y/F6pfWetSoUf7Y7SV9tcNOpUpuSpbyANCyVRYb1oeQ8zX//ns71Wu4iI72rrzzTifbtzlwOmHvHjs2Owx8OpJePfIxd3Yo7sBoBwH4bqeN8pVMipX0Ljdq6WbrBpsl/sPfG1Sp7qHQuc54rfoevt5u4HJCpVgPrTq5sdnAZoeyN5kkxvs+jr/r2512bqropngpb8BNHnCxeb3DEv+PB21Uq+Gm8Ln3/447XezYZsfphO59sujy7ywATiUZOJ0QmS9wrj39/isXJSvaKFLS2xzUbhHCt5+6LI2Ox3tokJHqXZeVDo6Q8485M0xcTnA5we0Ch7UfeU07+JWb0hVtRJ+Lv05LO19/6r5E/N5/s9LN7BiP/OChxr12bHYDR4jBzbXt7A6QjuBXO+xUvKDt+/QybV/9XNq+55+O5Mke+ZgXYG1fwq4zXH9jJPmLe7/UlLu/ML9uPm15702PCSY407yvkSvDjT3Ue/JNT3Jy9IsU6v7nBp/XPS/p2rUr69evv+iva9eulnJFixYlMTExezkhIYGYmJjs5TVr1nDixAkeeughnnjiCRISEujUqdNl93/Vc3j/+c9/Llq3YcMGkpOTARg5cuTVrsJlnThho0j0+Q9+dLRJWqpBWhrZKeLKld0sXxpKfLxB0aIma9eG4HQapKQYuN1Qo6aLHk9kkpUFQ/4bSWQ+aPtQlp8i+mtOnoBCOeIvFA3paQYZaWQPj5WvZLJumZ3EeDdFisKmdTZcToOzKRBb6/xzE+Nh7RI73foHzrejkwk2Csecj6FwtElamkF6GtnDYxUqe/hoaQgn4g2ii5p8utZxLn6D6wub2O0wcWQY2zbZqX2nmxKlAqcjlHzCpGCR89+qoooYZKZBZjqER3rXhUUYtIoLY9oz6URGGXg88MQYb1agxn0O9nzu4vUuqXjccFN1O5XvCJzhgeREk4LR5+MvWMT72c9Mg/Bz739YhEHbuBAmP51JZBSYHug91nvyLFPJxlfr3dxwiw2XE3b/z409QBKiiSds2R0cgCK5tH2VKrtZkaPtW7c2BJfT4My5tq96TRePn2v7XjzX9rUJkLYvLdFJRJGQ7OWIwiG40j240j3Zw2OOCDu3PVGSTYN/JLSAHdMDdw0v7y1fKIQ6A8v6pe5XnQ+bsAuHwC6lXr16vPHGGyQlJREREcG6desYPnx49uP9+vWjX79+ABw5coQuXbowf/78y273qmeErrvuOj777DMqV65M7dq1qV27NpGRkdn/vxb8/m3vQvYcr05sNTePds7k5aGRxD2ZD5sBBQp4CHFA8xZOesdlEhoK+fND23ZZbPk8cE4EuWQbAbDliL9yNZM2j7qZ+JKDoX0cGAbkK2BiP9+GcOigwYinQ7ivtZvqdQKnI+D5E/HfUs1D+y7/3959h0VxvW8Dv7fSBOkIamKJEWPEjmIBsSv28jUxEcWSYI01GmPEHluCPYlGE40aW2KNDVtsWGJsv9hrRFCqAtJ39/1jeVdHUIkuMyxzf65rr4tZzu48h5k9+/CcM7vZmB1mhc8HWUOhAErYG6BWP33wZ19kYsXvaUhNBjat1uTzjEVTQY7/g9s6HFybhelfBI4AACAASURBVGE/2GLsajs06aHBr9MzYDAYcGBtFmwdFBi3xg6fr7JDeipw9HfLeCMEjElNfpTPJDMxt/XYtzYHo3+wwldrbND0Aw1+mZYFg8GAdgM0gAKYNyQTq6ZmoVJNJVQW8vIv6Nj3Ua9MTJ1oi2HPjH1qNdAmKBsDnxn7OlvY2PeiN3uF8mli/PhuBq5sjEWz8Epos7QKKndxw8lv/s13moYKl4eHB0aMGIHg4GB06tQJ7dq1g4+PDwYMGICLFy++9vMW+hk7duxY+Pv7Y968eRg5ciTq1auHlStXonPnzoW96wJzdzfgypWnJ358vAIl7A2wtnnaJi0NqFY9B63bGtcFJCUqsPInK9g7GLAvQoMKFXSoUDF3VDHAYgZCAHBxA24+0/+keGOSY/VM/9PTAG8fPQLaGPv4OAn4baUKJeyNvz9xUImVi1ToNViHBk1fMLoWUW7uely//PSAJeZz/NPTgPd8dGjWxljpepRkXBtUwgE4d1qFt8rr4exqgI0N0KipDieOWEhJAICjmwJRV58O6snxBtiUALTWT8+JG3/r8NZ7Krh4Gt8h67XTYOeyLKQlA5eO69AuVAu1RgG1BqjZzFghatRF9K68Fkd3Bf69+vScza//184YKz4uXsb+N2inwval2UhLBrIyDQjqp4GtvbH9wQ3ZcPGyjIXi7u4GXC3g2NfqmbFvVe7Ytz937CufO/YZDIDagsY+G1cNEq+nmbYzErOhsVNBbf00E4w9nwJnb1vT4ugKrVxw4ecYZKXouDBaAu3bt0f79u0F9y1btixPuzJlyuDAgQMFek5R1gj5+fnhhx9+wNq1azFr1izoitgkcu06ObhySYX7UcY/xx/btfBrIFwsmZCgxOcj7fAkd43AmtVWaNI0BwoFcOe2Eqt+toJOB2RmAtu2ahHQxHKuGqpWW4+blxV4cN+4fWCHCrX8hMnMowRgxhgN0nP7v3WNCn6BeigUwKnDCvyyRIUxX+dYXBIEANVr63D9sgoxUcY3hL3b1ajbQDi1l5igQNgoa6Tl9n/Tai0a5R7/43+qsPEX47qK7Czj9vs1itY5/jLv1FLh3hU94u8bj93pndnwri8c4D0rKnHnog6pScY2lyN1cPJQwK6kAl4Vlfi/I8a/ly7HgCsnc1DG23ISwXdrqfDvFT3icvt/YqcOVf2E8Zd+R4lbF/VISTImjP9E6uGc2/8TO3XY+4vx9Z6SZMCp3TrUbGIZb5C1nhv7duYz9iU+N/b9+tzY98szY9/2rVr4W9DY51HdHknX05EaY1z4f3tvIjzrCqdoHMvbIOHSE2Q8MvYr+nQy7Ny1xT8JKoKXzxcWhUHk+t7GjRuxa9curFix4rWf406UpxkjMjp1Uo0VP1ohJwfw9NRjzLh0PIhRIvwbG3y31DgCbN2iwfatWhj0QNX3dRg8LANWVkBGBrB4oTWuXFIhRwc09s9BSL9MFMbVww911q9u9BrOn1JgwwoVcrIVcPcy4NMxOYh9oMCKb1WY9r3xTS5iq/HyeYMBeLeqHsG5l8iP6aNB2hPAyeXpqVSpqgG9h5o/GbBVFs4g+/dJFdYs1yAnRwEPTz2Gjs3Ewxglvv9Wi7k/ZAAAdm1RY/c2DQx6wPt9HfoNzYKVlfHy+R/mWeHeHSUUAOo2zEGP3tmCqSVzuZLlYf4nBXD1dA4ifs6CLscA51JKdB1tjaQYPTYvyMSQRcaFQie2Z+Hkjmyo1ArY2CvQbqAWHm+rkJZsvHw++qYOSqUCFWqo0Ka/Fiq1eV8AWkXhJZeXT+mw++ds6HIAZ08FPhitRUKMAZvmZ2HEYuNr7vj2HBzfngOVGrCxV6DTIA1Kva1ERpoB6+dmIT7auKg2sIcatZqa/02ymrZwrkA4dVKNn58Z+0aPS0dMjBLzv7HB4tyxb9sWDXZs1UKfO/YNembsW5I79ulyx77ehTT2LU1saP4nBfDg79zL53MMsPPQos6QMngSm4Wz391H07mVAAC3difg5u4EKNUKaEuoUL2fFxzKCsfizd0vou3yKoWWIM302VQoz/si1YeFi7av8wtGiLav/IieCJlDYSRClqKwEiFLUViJkKUorETIEhRmImQJCisRshSFlQhZCrEToRpDxUuEzi2UNhGynA87ISIiIjKzYj7JSURERP+Zxc0VvT5WhIiIiEi2WBEiIiIiAQUrQkRERETFHytCREREJMSKEBEREVHxx4oQERERCXCNEBEREZEMsCJEREREQqwIERERERV/rAgRERGRECtCRERERMUfEyEiIiKSLU6NERERkQAvnyciIiKSAVaEiIiISIgVISIiIqLijxUhIiIiElAY5FMSYkWIiIiIZIsVISIiIhKST0GIFSEiIiKSL1aEiIiISICfI0REREQkA6wIERERkRArQkRERETFHytCREREJMA1QkREREQywIoQERERCbEiRERERFT8MREiIiIi2eLUGBEREQlwsTQRERGRDFhkRSi03QCpQyCJZLnZSh2CpDRJGVKHICnFrftShyAdg4z+Rae8kkTen4xON1aEiMgiyDoJIqJCY5EVISIiIio8XCNEREREJAOsCBEREZGQjNaksSJEREREssWKEBEREQlwjRARERGRDLAiREREREKsCBEREREVf6wIERERkYBCL3UE4mFFiIiIiGSLFSEiIiIS4hohIiIiouKPiRARERHJFqfGiIiISIAfqEhEREQkA6wIERERkRC/dJWIiIio+GNFiIiIiAS4RoiIiIhIBlgRIiIiIiFWhIiIiIiKP1aEiIiISIBrhIiIiIhkgBUhIiIiEuLnCBEREREVf6wIERERkQDXCBERERHJACtCREREJMSKEBEREVHxx0SIiIiIZItTY0RERCTAxdJEREREMsCKEBEREQnp5VMSYiKUy7fxuwgZ2hwarRq3rz9A+KStSHuSmW/bUVM6486NWPy26hgAQKlUYNC4IFSrXQ4AcProdfwYvkes0M3iTfpfwsEGQ79sh4qVPZGRnoW9W89i27qTYob/xur5VkT/fgHQalS4dTsOc77ZibS0LEGb5s2qokf3ejDAgMyMbCxcsg/Xrj0AAHy/uA+0Vmrk5OgAAPv3/4P1G0+J3g9z8G1YCSFDmkOjVeH29YcIn7otz7nQtI0PuvdqAIMByMzIxpK5u3D9crREEZuXb4tqCJnYxfhauBSF8GErkZaSkW/bUYtCcOfKffy2aK/IUb4Z35bVEDKxq7GP/0QhfNjPefr4ojYlHO0w9JuPUbFaWWSkZWLvmmPYtuwA3qrsibHLBpger1QpUf69MpjaawmO7fhb7C6+lNz7T0KcGgNQ0skWIyd3wtTR69C/0wLERCUh5LMWedqVLe+KmUv7oHGLqoL7m7WrjjLlXDGw+2IM6rEEPnXK5WlTlL1p/z8d0xoZaVn4pMtCDO+1DHUaVYJv43fFCv+NlSxpg89Ht8WkKZvRu+8yRMc8woB+TQRtypZxxqcDAjF2/Hp8EvoTVq89jslhnQEA1tYaeHk5YsCnK/BJ6E/4JPQni02CSjraYlRYJ0z9fD36d12EB/eT0HdIc0GbMm+7oP9nLfDl0NUY9NH3WLviMCbO6SFRxOZV0qUERi7qg6m9v0P/el8h5k48QiZ2ydOu7LulMHPLKDTuVFuCKN+MsY8hmBq8BP19JyDmbhxCwroWuM2nM3og40kGPqn/FYa3mIE6LarBt5UP/r0ag8H+U0y3vw9cwsFNJ4tcEiD3/heYQcSbxERJhC5cuGD6OTIyEjNnzsTcuXNx/vx5MXb/SrX83sG1f6IR/W8iAOCPjafRtI1Pnnbte9RDxNazOBLxj+B+pVIJaxstNFo1NBo11GoVsjJzRIndHN60/5WqeGH/H+eh1xuQk6PD6SPXLCoRrFO7PK5ei8H9+0kAgG3bz6JZs/cEbbKydZj77S4kJj4BAFy99gDOTiWgVivhXdkT6elZ+Hp6d/y4tC8GhTaDVmuZxdZa9Svi6qX7iL5nPBd2bPoLTdtUE7TJztJh3tRtSExIBQBcuxQNJ5cSUKtVosdrbrUCq+La2TuIvhULAPhjxSE07V4vT7v2/QIRsfYYjmw5I3aIb6xW0+f6uDxvH1/WplKNt7F//Qnj6z1bh9N7L6BxB2FCWNWvEhp1rI2FI38RoUf/jdz7T3mJkgiFhYUBANasWYMZM2agVKlScHV1xcSJE7F69WoxQngpN4+SiHvw2LQd9zAZdvbWsLWzErRbMvMP7P8jb/IWse0sUpPTsXrvaKzdNwbR9xJw8vDVQo/bXN60/1cvRqFZUHWo1MaEsGGz9+Dsal/ocZuLu5sDYuNSTNtxcckoYWcNW1ut6b6HDx/j5Kmbpu2BnzbF8cjryMnRw9ZWi3Pn/8WkKZsxcPBKuLs7YEC/AFH7YC5uHiUR/zDZtB0Xmwy7EsJz4WHMI5w6dt20/emIVjhx+KppWtCSuZV2QlxuQgwAcdFJsHOwha29taDdkrG/Yv+GE2KHZxZupZ0Rdz/RtJ1fH1/W5upft9CsR32o1CpY21mhYftacPYoKdjHgCndsXLa5hdOKUpJ7v0vKIVBvJvURJ0a27BhA1atWoU+ffqgT58+WLNmTZFIhBRKRb7363T6Aj3+o08D8TjpCT5sOhsft5oL+5K26NKrgTlDLFRv2v+l3+6BwQAsXjcQE8M/xNkTN5GdbTlvii/qvz6fxYLW1hqEfdUJpUs7Ye63uwAAxyNv4OtZO5CWloXsbB3W/HocjRpaztTgs5T/4Vywstbgy5nd4VXWGeFTtxV2aKJQKPMfEgv6WrAEBXm9v6zN0gkbjK/3wxMx8ZfBOHvokuD1XsW3IhxcSuDgxqK5TlDu/ae8REmEcnJyoNfr4eLiAltbW9P9Wq0WyhcMPGKKi3kEZ9cSpm1Xd3ukPE5DZkZ2gR7fsFkV7NlyFjk5OqSlZmLf9rOoXrd8YYVrdm/af1s7K/w4by9Cuy3G+NCV0OsNiL6XUFjhml1sbDJcnJ/2383VHsnJ6ch4rv/ubg5YOK8XdDo9Ro7+FU9yFxD71X8HPtXKmtopFArk5FjmG2fsg8fCc8HNHimP0/OcC24eJTFvRT/o9QZ8HvoznqRa7n++z4qLShD8d+/q6YiUpCfIfG7hvCWLi0oU9tErbx9f1sbWwQY/hm1CaIMwjO/yrfH1njuFBAABneti/7pIGAxF4F/9fMi9/wVmMIh3k5goWYiTkxMCAgJw48YN0zRZZGQkPvjgA7Ru3VqMEF7qTORNePuUhddbzgCAoG51EXnoSoEff+NyDPxbGtfEqNRK1A/wxpULUYUSa2F40/4Hda+L4EFNAQCOznZo06U2Du268IpHFR1/nbmNKlW8ULq0EwCgfbuaOB55XdDG3t4a4d/0xJGjVzFtxjZkZT1dA+bmao/QTwKh1aqhVCrQvWtdHPzzsqh9MJczJ27C+/0y8Cqbey50rYPIP4Xngr2DDeYu7YOjBy/j6/GbLGo93KucOXgJ3nUqwKuCOwAgKCQAkbvOSRyVeZ058A+861R8po9NELnzXIHbBIUEIHh8RwCAo5sD2gT749Cmp9WPag3fxbnDRff8l3v/KS+FQcS09datW0hOTkaNGjVw5swZpKSkoEmTJv/5eVrXmGj22Oo2qoSQoS2g1qgQE5WIORN+h2cZJwwP64jBPb4TtH3+8nH7kjYYNC4I73h7Qq834NzJW1j67W7oLKgq8Cb9t7HVYsz0rvAq6wyFQoH1yw/jwM7CSYSy3Gxf3eg11POtgP59m0CtUSI6+hFmzt4BT09HjB7ZBp+E/oSPevqhT3Bj3L4TJ3jc6DG/IiU1A5/0D0T9+hWhUilx7ty/WLg4olCmBzVJhV95qduwEvoObpZ7LiRhTthmlCrthBETOmDQR9/jw76N0evTQNy5ESt43NhBK5HyOL3Q4lLcul9oz/2sus3fR8jELlBr1Yi5HYc5A5fDs5wbhs/vjcEBUwRtRb183oxDdd3cjwhQa9SIuROLOaEr4FnOFcMX9MZg/ykvbJP66AlsSlhhzPf94VXBHQoA6+ftwoFn1kttiVqM/r4TEB+d9IK9S88S+7876UezPt+rBLaaJdq+Du4ZK9q+8iNqImQuhZEIkWUorETIUoiRCBVVYiVCRZblDdVkRkyECo9lXuNLREREhUdGebf0K5WJiIiIJMKKEBEREQkoZDQVy4oQERERyRYTISIiIpItTo0RERGRkOV8+ssbY0WIiIiIZIsVISIiIhLgYmkiIiIiGWBFiIiIiITkUxBiRYiIiIjkixUhIiIiEuIaISIiIqLijxUhIiIiElDIpyDEihARERHJFytCREREJMQ1QkRERETFHytCREREJKDgd40RERERFX+sCBEREZEQ1wgRERERFX+sCBEREZGQfApCrAgRERGRfDERIiIiItni1BgREREJKLhYmoiIiKj4Y0WIiIiIhFgRIiIiIipatm/fjrZt26JFixZYs2ZNnt/v27cPHTt2RIcOHTBo0CA8fvz4lc/JRIiIiIiE9CLeCujhw4cIDw/H2rVrsXXrVqxfvx43btww/T41NRWTJk3C0qVLsW3bNlSuXBkLFy585fMyESIiIiLJJCcnIyoqKs8tOTlZ0O748eOoX78+HB0dYWtri1atWmH37t2m32dnZ2PSpEnw8PAAAFSuXBkxMTGv3D/XCBEREZGAmFeNrVy5EosWLcpz/5AhQzB06FDTdmxsLNzc3Ezb7u7uuHDhgmnbyckJzZs3BwBkZGRg6dKl6NWr1yv3z0SIiIiIJNO7d2907tw5z/0ODg6CbUM+yZlCochzX0pKCgYNGgRvb+98n/d5TISIiIhISMSKkIODQ56kJz8eHh7466+/TNuxsbFwd3cXtImNjUW/fv1Qv359jB8/vkD75xohIiIiKvIaNGiAyMhIJCYmIj09HXv37oW/v7/p9zqdDqGhoWjTpg2+/PLLfKtF+bHIipAiJk7qECRjyM6ROgRJWaU6Sh2CpAwJSVKHIJ0CDmrFlSE7W+oQJKVPS5M6BHkpgp8j5OHhgREjRiA4OBjZ2dno1q0bfHx8MGDAAAwbNgwPHjzApUuXoNPpsGfPHgDA+++/j+nTp7/0eRWG/Cbdirg2HgOlDkEyck+EFM5MhGRL7olQVpbUIUhK7olQhH6jqPtrVTtMtH3tOTNZtH3lxyIrQkRERFSI/sPn+1g6rhEiIiIi2WJFiIiIiAT47fNEREREMsBEiIiIiGSLU2NEREQkxKkxIiIiouKPFSEiIiISYkWIiIiIqPhjRYiIiIiEWBEiIiIiKv5YESIiIiIhfsUGERERUfHHihAREREJ8Cs2iIiIiGSAFSEiIiISYkWIiIiIqPhjRYiIiIiE9KwIERERERV7rAgRERGRENcIERERERV/TISIiIhItjg1RkREREKcGiMiIiIq/lgRIiIiIiFWhIiIiIiKP1aEiIiISIgfqEhERERU/LEiREREREIGvdQRiIYVISIiIpItVoSIiIhIiFeNERERERV/rAgRERGRkIyuGmMilKtu8/cR8mVHaLQa3L4UhXkjViMtNSPftiPnB+PulWj89t0+wf2uXk4I3/k5BjedhuTEJ2KEXWh8W1RDyMQu0GjVuH0pCuHDViItJf+/x6hFIbhz5T5+W7RX5CjNp24Tb4SMbguNVoXbV2Mw74uNSEvNzLftyFk9cPfaA/y2/E/TfetOhSH+QbJp+7cfD+HgtrOFHvfr8m1ZDSETuxqP7z9RCB/2c57j+6I2JRztMPSbj1GxWllkpGVi75pj2LbsAN6q7ImxywaYHq9UKVH+vTKY2msJju34W+wuvjY5nPu+raojZHL33GN7D+GDluc9/i9oo1QqMPjbYFRrVBkAcHrPBSz7ch0AwN7JDoPmfoy3vEvDykaDX2dvx/51x0Xv36v4tq2FfjN6QmOlwe0Ld/FN/++QlpJe4DYbHy5Hwv1EU9sNc7fiwNqjcPF0wugVg+BUyhFKpRLrZ2/B/jVHRO0b/XecGgNQ0qUERs4PxrS+SzGg4SQ8uBuPkAmd8rQrW6kUvv5tOBp3qJ3nd82618PcraPg6ukoRsiFqqRLCYxc1AdTe3+H/vW+QsydeIRM7JKnXdl3S2HmllFo3Cnv38OSlHS2w8hZPTBt8CoMaDkHD/5NRMiYtnnala3ojq9/+RSN2/oI7i9d3g0pj9MxpEO46VaUkyDj8Q3B1OAl6O87ATF34xAS1rXAbT6d0QMZTzLwSf2vMLzFDNRpUQ2+rXzw79UYDPafYrr9feASDm46aVFJkBzO/ZKu9hj1fX9M/Wgh+tcahwe349B3yv8K3KbZhw1RplIphPp+iYH1v0K1RpXRuHNdAMCo7wcg/n4SBjeciHHtZmPgnI/h6uUkeh9fpqSrA0avGIQp3eaib5XPEHP7IfrN/KjAbcq864XUpFSE1hpjuh1YexQA0Hd6T1w5dQOhNcfgizbTMWzJADh5WOh7gsEg3k1ioiVCR44cQXKy8T/mLVu2YMqUKfjtt9/E2v1L1WpSBdfO3kH07TgAwI6VhxHY1TdPu3YhAYj49TiObDsjuN/ZoyT82lTHxI8WiRJvYasVWNX497gVCwD4Y8UhNO1eL0+79v0CEbH2GI5sOZPnd5akVqN3ce3CPUTfjQcA7FgbicAONfO0a/dxA0T8dhpHdl4Q3P9erbeh1+kxc/WnWLJjJHoOaQ6lUiFK7K+jVtPnju/yvMf3ZW0q1Xgb+9efgF5vQE62Dqf3Xsjzz0FVv0po1LE2Fo78RYQemY8czv1aTd/H1TO3EH3zIQBgx48H0PR/fgVuo1QpYW1rBY2VBhorNTRaNbIysmHvZIdaTati9ddbAADx0Un4LHAyUpKKVnW8dksfXDt9E/dvPAAAbP9uL5r1bFzgNlUbvAu9To85+8Pww7m5+PirblAqjW+lSpUSdiVtAQDWtlrocnQw6OVzGbqlEmVqbPr06bh8+TLCw8Mxb948XLx4Ec2aNUNERAQuX76MCRMmiBHGC7l6OSEuOsm0HR/9CHYONrAtYS2YHvtu/HoAQI3G3oLHJz58jGl9l4oTrAjcSjsh7v7Tv0dcdBLsHGxha28tKJ8vGfsrAKCGfxXRYzQnV09HxMU8Mm3HP3gMO3sb2JawEkyPfTfZOMDX8KskeLxKrcTZY9exfNYOaK00mPxjP6SlZmDLz0fF6cB/5FbaGXHPlPXzO74va3P1r1to1qM+/jl5AxorNRq2rwVdtk6wjwFTumPltM0vnFIqquRw7ruVcUb8s8f2fiLsSj53/F/SJmL1Efh3ros11+ZBpVbi7wP/h5O7zqFy7QpIfPAIXYa2Rt0WPtBYqbFpwS7cv/FQ9D6+jFtZV8RFxZu246IScvtmY5r6elkbpVqFM/suYNmYX6C10WL6ji/wJDkNm+fvxPLxa/Dt4anw71YfJd0c8MPoVXgUl5wnBotQBCo1YhElETp27Bi2b98OlUqFQ4cOYcOGDdBqtejRowfatWsnRggv9f+z+efpZJrJK17099AVz7/Hi6o3Be3v7vWnTD9nZ+mwecVhdOzdsMgmQooC9PdlbZZO2IABU/+HxYcnIvHBY5w9dAlVfN8xtaniWxEOLiVwcONJ8wYuAjmc+wU531/W5uPxnfAoPgUfVBgKrY0Wk9Z9hq5DW+Py6ZvwLO+OtJR0jGwxDV4V3DF375e4f+Mhbpy7UxhdeS0v6pu+AP3X6/TY9eN+03Z2Vg42he9A56FtsXn+Tnyx+jNsmLMVO77fi9LvlMLcg5Nx+cR1XD19w7ydILMSZWrM2toaCQkJAAAXFxekpaUBANLT06FWS79eOzYqEc4eJU3brp6OSEl6gsy0LAmjkk5cVIKs/h6x0Y/g7O5g2nb1cEDKozRkpmcX6PFNO9VCucqepm2FAsjJLrpvnHHPn+9eeY/vy9rYOtjgx7BNCG0QhvFdvoVebzBNJQFAQOe62L8uEgYL/I9SDud+7L1EOD+zbsXVywkpiamCPr6sTcMOdbD3l8PIydYhLTkdEWuOorp/FSTEGCtpEauNi4Ojb8Xin8hr8K5TQaSeFUzsv/FwLvV03ZJraWckJ6YiIy2zQG2af+yP8tXeMv1OoVAgJzsHDi72qNrIGzuXGS+iuX/jAf6OuIBqFlg1lBtREqEhQ4agW7dumDVrFipUqIBevXphxowZ+N///oeQkBAxQnipv/+8DO/a5eFV3g0A0LZ3Y0TuPi9xVNI5c/ASvOtUgFcFdwBAUEgAInedkziqwvP3kavwrvEWvN52BQC07emHyH3/FPjx5SqVQq/hLaFUKqC1UqN9r4Y4vLPo/r3OHPgH3nUqPnN8myDyuXhf1iYoJADB4zsCABzdHNAm2B+HNj2t/lRr+C7OHb4sRlfMTg7n/pkDF+HtWxFeFT0AAEH9miLyj7MFbnPj3F34dzGum1KpVagfVBOXT9/Ew7vxuH72Dpp/1AgA4OjugPfqVcK1v2+L1bUCObP3PKrUr4TS75QCALQLbYnIracL3Kbc+2XRe3IPKJVKaK216Di4NQ5tOI7khBTERyXAv1t9AICDiz2q+VfBlZPXReydGclosbTCINK/bffu3cO+fftw9+5d6HQ6uLq6IjAwED4+Pq9+8HPaeAw0e3x1m1VFny87Qa1RIeZuPOYO+Rmeb7vis28/xpBmMwRtX3T5PADsevgdelQZXWiXzxuycwrleZ9Xt/n7CJnYBWqtGjG34zBn4HJ4lnPD8Pm9MThgiqCtmJcQK5wL5wqMugHe6DO6jfH4/5uAuWPWwbOsCz6b0R1DOoQL2j5/+byVtQaDwjrBu+bbUKmVOLLrAlZ+s7tQ4jQkJL26UQHUzb1EXK1RI+ZOLOaEroBnOVcMX9Abg/2nvLBN6qMnsClhhTHf94dXBXcoAKyftwsHNpwwPfeWqMXo7zsB8dHmidVEIc4C9KJ67huyzFeVqtvSB30nKaQlwwAAB+1JREFUdzf28VYs5nyyFKXKuWHE4r4Y1GDiC9ukJD2BvbMdBs/thXdqvA29zoCzh/7B0i/WQZejg1sZZwz5Nhie5d2hUCqwefEe7FxxyCwx63NnEszBt01N9J3RExqtGtE3H2J270XwrOCOkcsGIrTWmBe2SUlKhZWNFkMW9UOVeu9CrVHh8KZIrPjSuGasgs/bGLygLxxc7GHQG7B54U7BVNqbiNBvNMvzFFSb0kNF29eu+wtF21d+REuEzKkwEiFLIVYiVFQVViJkKcyVCFkkkRKhosqciZAlMmciZIlET4Q8B4u2r10xi0XbV374OUJEREQkW9KvVCYiIqKixfImi14bK0JEREQkW6wIERERkRArQkRERETFHytCREREJKRnRYiIiIio2GNFiIiIiAQMhqL7NUHmxooQERERyRYrQkRERCTENUJERERExR8rQkRERCTEzxEiIiIiKv6YCBEREZFscWqMiIiIhPS8fJ6IiIio2GNFiIiIiIS4WJqIiIio+GNFiIiIiAQMXCNEREREVPyxIkRERERCXCNEREREVPyxIkRERERC/NJVIiIiouKPFSEiIiISMvCqMSIiIqJijxUhIiIiEjBwjRARERFR8ceKEBEREQlxjRARERFR8cdEiIiIiGSLU2NEREQkwMXSRERERDLAihAREREJyWixtMJgkNFXzBIRERE9g1NjREREJFtMhIiIiEi2mAgRERGRbDERIiIiItliIkRERESyxUSIiIiIZIuJEBEREckWEyEiIiKSLSZCREREJFtMhP6D7du3o23btmjRogXWrFkjdTiiS01NRbt27RAVFSV1KKJbtGgRgoKCEBQUhNmzZ0sdjujmz5+Ptm3bIigoCD/99JPU4Uhm1qxZGDdunNRhiC44OBhBQUHo2LEjOnbsiPPnz0sdkqgOHDiALl26oHXr1pg2bZrU4ZCZ8bvGCujhw4cIDw/H77//Dq1Wiw8++AD16tXDO++8I3Voojh//jwmTJiAO3fuSB2K6I4fP46jR49i8+bNUCgU6N+/PyIiItCiRQupQxPFqVOncOLECWzbtg05OTlo27YtAgICUKFCBalDE1VkZCQ2b96MJk2aSB2KqAwGA27duoVDhw5BrZbfW8a9e/cQFhaGjRs3wsXFBb1798aff/6JgIAAqUMjM2FFqICOHz+O+vXrw9HREba2tmjVqhV2794tdVii2bBhA8LCwuDu7i51KKJzc3PDuHHjoNVqodFoULFiRURHR0sdlmh8fX2xatUqqNVqJCQkQKfTwdbWVuqwRPXo0SOEh4cjNDRU6lBEd+vWLSgUCgwYMAAdOnTA6tWrpQ5JVBEREWjbti1KlSoFjUaD8PBwVK9eXeqwyIzkl96/ptjYWLi5uZm23d3dceHCBQkjEtf06dOlDkEylSpVMv18584d7Ny5E+vWrZMwIvFpNBosWLAAK1asQOvWreHh4SF1SKKaOHEiRowYgZiYGKlDEV1ycjL8/PwwadIkZGRkIDg4GOXLl0fDhg2lDk0Ud+/ehUajQb9+/RAXF4fAwEAMHz5c6rDIjFgRKiCDwZDnPoVCIUEkJJXr16+jb9++GDt2LMqVKyd1OKIbNmwYIiMjERMTgw0bNkgdjmg2btwIT09P+Pn5SR2KJGrWrInZs2fD1tYWzs7O6NatG/7880+pwxKNTqdDZGQk5syZgw0bNuDixYvYvHmz1GGRGTERKiAPDw/Ex8ebtmNjY2U5TSRXZ86cQZ8+fTBq1Ch07txZ6nBEdfPmTVy+fBkAYGNjg5YtW+Lq1asSRyWenTt34tixY+jYsSMWLFiAAwcOYMaMGVKHJZq//voLkZGRpm2DwSCrtUKurq7w8/ODs7MzrK2t0axZM1nNBsgBE6ECatCgASIjI5GYmIj09HTs3bsX/v7+UodFIoiJicHgwYMxd+5cBAUFSR2O6KKiojBhwgRkZWUhKysL+/fvR+3ataUOSzQ//fQTduzYga1bt2LYsGFo2rQpxo8fL3VYoklJScHs2bORmZmJ1NRUbN68WTYXCgBAYGAgjh49iuTkZOh0Ohw5cgRVq1aVOiwyI/mk9W/Iw8MDI0aMQHBwMLKzs9GtWzf4+PhIHRaJYPny5cjMzMTMmTNN933wwQf48MMPJYxKPAEBATh//jw6deoElUqFli1byjIhlKvAwEDT8dfr9ejZsydq1qwpdViiqV69Ovr374+ePXsiOzsbDRs2RNeuXaUOi8xIYchv8QsRERGRDHBqjIiIiGSLiRARERHJFhMhIiIiki0mQkRERCRbTISIiIhItpgIERERkWwxESIiIiLZYiJERAXSpk0b+Pv74/r161KHQkRkNkyEiKhAduzYgXLlymHPnj1Sh0JEZDZMhIioQFQqFWrXri2rL1wlouKP3zVGRAWSkZGBP/74A/xWHiIqTlgRIqICCQ8Ph4eHB+7du4cnT55IHQ4RkVkwESKiVzp79ix2796NhQsXwt7eHteuXZM6JCIis2AiREQvlZmZiS+++AKTJ0+Go6MjvL29uU6IiIoNJkJE9FLz589HzZo10aRJEwCAt7c3rly5Im1QRERmwkSIiF7owoUL2L17N8aPH2+6r0qVKqwIEVGxoTDwEhAiIiKSKVaEiIiISLaYCBEREZFsMREiIiIi2WIiRERERLLFRIiIiIhki4kQERERyRYTISIiIpItJkJEREQkW/8PQn7TeqHNlUwAAAAASUVORK5CYII=\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    },
-    {
-     "data": {
-      "image/png": "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\n",
-      "text/plain": [
-       "
    "
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    }
-   ],
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
    "source": [
     "# optional\n",
     "# visual representation of grid search\n",
@@ -3319,7 +2557,9 @@
   {
    "cell_type": "code",
    "execution_count": 13,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "pip3 install tensorflow"
@@ -3335,7 +2575,9 @@
   {
    "cell_type": "code",
    "execution_count": 14,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "conda install tensorflow"
@@ -3351,7 +2593,9 @@
   {
    "cell_type": "code",
    "execution_count": 15,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "# import necessary packages\n",
@@ -3401,7 +2645,9 @@
   {
    "cell_type": "code",
    "execution_count": 16,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "from keras.utils import to_categorical\n",
@@ -3431,7 +2677,9 @@
   {
    "cell_type": "code",
    "execution_count": 17,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "import tensorflow as tf\n",
@@ -3577,7 +2825,9 @@
   {
    "cell_type": "code",
    "execution_count": 18,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "epochs = 100\n",
@@ -3592,7 +2842,9 @@
   {
    "cell_type": "code",
    "execution_count": 19,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "DNN_tf = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)\n",
@@ -3615,7 +2867,9 @@
   {
    "cell_type": "code",
    "execution_count": 20,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "# optional\n",
@@ -3654,7 +2908,9 @@
   {
    "cell_type": "code",
    "execution_count": 21,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "# optional\n",
@@ -3678,7 +2934,9 @@
   {
    "cell_type": "code",
    "execution_count": 22,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "conda install keras"
@@ -3694,7 +2952,9 @@
   {
    "cell_type": "code",
    "execution_count": 23,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "pip3 install keras"
@@ -3710,7 +2970,9 @@
   {
    "cell_type": "code",
    "execution_count": 24,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "from keras.models import Sequential\n",
@@ -3733,7 +2995,9 @@
   {
    "cell_type": "code",
    "execution_count": 25,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "DNN_keras = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)\n",
@@ -3756,7 +3020,9 @@
   {
    "cell_type": "code",
    "execution_count": 26,
-   "metadata": {},
+   "metadata": {
+    "collapsed": false
+   },
    "outputs": [],
    "source": [
     "# optional\n",
@@ -3799,73 +3065,103 @@
     "\n",
     "## Which activation function should I use?\n",
     "\n",
-    "Backpropagation algorithm works by going from the output layer to the\n",
-    "input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the\n",
-    "cost function with regards to each parameter in the network, it uses these gradients to update each\n",
-    "parameter with a Gradient Descent step.\n",
+    "The Back propagation algorithm we derived above works by going from\n",
+    "the output layer to the input layer, propagating the error gradient on\n",
+    "the way. Once the algorithm has computed the gradient of the cost\n",
+    "function with regards to each parameter in the network, it uses these\n",
+    "gradients to update each parameter with a Gradient Descent (GD) step.\n",
     "\n",
-    "Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower\n",
-    "layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually\n",
-    "unchanged, and training never converges to a good solution. This is called the vanishing gradients\n",
-    "problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many\n",
-    "layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients\n",
-    "problem, which is mostly encountered in recurrent neural networks. More generally,\n",
-    "deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds\n",
+    "\n",
+    "Unfortunately for us, the gradients often get smaller and smaller as the\n",
+    "algorithm progresses down to the first hidden layers. As a result, the\n",
+    "GD update leaves the lower layer connection weights\n",
+    "virtually unchanged, and training never converges to a good\n",
+    "solution. This is known in the literature as \n",
+    "**the vanishing gradients problem**. \n",
+    "\n",
+    "In other cases, the opposite can happen, namely the the gradients can grow bigger and\n",
+    "bigger. The result is that many of the layers get large updates of the \n",
+    "weights the\n",
+    "algorithm diverges. This is the **exploding gradients problem**, which is\n",
+    "mostly encountered in recurrent neural networks. More generally, deep\n",
+    "neural networks suffer from unstable gradients, different layers may\n",
+    "learn at widely different speeds\n",
     "\n",
     "\n",
     "## Is the Logistic activation function (Sigmoid)  our choice?\n",
     "\n",
-    "Although this unfortunate behavior has been empirically observed for quite a while (it was one of the\n",
-    "reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that\n",
-    "significant progress was made in understanding it. \n",
+    "Although this unfortunate behavior has been empirically observed for\n",
+    "quite a while (it was one of the reasons why deep neural networks were\n",
+    "mostly abandoned for a long time), it is only around 2010 that\n",
+    "significant progress was made in understanding it.\n",
     "\n",
-    "A paper titled **Understanding the Difficulty of Training Deep Feedforward Neural Networks** by Xavier Glorot and Yoshua Bengio1 found a few suspects,\n",
-    "including the combination of the popular logistic sigmoid activation function and the weight initialization\n",
-    "technique that was most popular at the time, namely random initialization using a normal distribution with\n",
-    "a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this\n",
-    "initialization scheme, the variance of the outputs of each layer is much greater than the variance of its\n",
-    "inputs. Going forward in the network, the variance keeps increasing after each layer until the activation\n",
-    "function saturates at the top layers. This is actually made worse by the fact that the logistic function has a\n",
-    "mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the\n",
-    "logistic function in deep networks).\n",
+    "A paper titled [Understanding the Difficulty of Training Deep\n",
+    "Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio](http://proceedings.mlr.press/v9/glorot10a.html) found that\n",
+    "the problems with the popular logistic\n",
+    "sigmoid activation function and the weight initialization technique\n",
+    "that was most popular at the time, namely random initialization using\n",
+    "a normal distribution with a mean of 0 and a standard deviation of\n",
+    "1. \n",
+    "\n",
+    "They showed that with this activation function and this\n",
+    "initialization scheme, the variance of the outputs of each layer is\n",
+    "much greater than the variance of its inputs. Going forward in the\n",
+    "network, the variance keeps increasing after each layer until the\n",
+    "activation function saturates at the top layers. This is actually made\n",
+    "worse by the fact that the logistic function has a mean of 0.5, not 0\n",
+    "(the hyperbolic tangent function has a mean of 0 and behaves slightly\n",
+    "better than the logistic function in deep networks).\n",
     "\n",
     "\n",
     "## The derivative of the Logistic funtion\n",
     "\n",
     "Looking at the logistic activation function, when inputs become large\n",
-    "(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when\n",
-    "backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what\n",
-    "little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so\n",
-    "there is really nothing left for the lower layers.\n",
+    "(negative or positive), the function saturates at 0 or 1, with a\n",
+    "derivative extremely close to 0. Thus when backpropagation kicks in,\n",
+    "it has virtually no gradient to propagate back through the network,\n",
+    "and what little gradient exists keeps getting diluted as\n",
+    "backpropagation progresses down through the top layers, so there is\n",
+    "really nothing left for the lower layers.\n",
     "\n",
-    "In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the\n",
-    "signal to flow properly in both directions: in the forward direction when making predictions, and in the\n",
-    "reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it\n",
-    "to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the\n",
-    "outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have\n",
-    "equal variance before and after flowing through a layer in the reverse direction (please check out the\n",
-    "paper if you are interested in the mathematical details). \n",
+    "In their paper, Glorot and Bengio propose a way to significantly\n",
+    "alleviate this problem. We need the signal to flow properly in both\n",
+    "directions: in the forward direction when making predictions, and in\n",
+    "the reverse direction when backpropagating gradients. We don’t want\n",
+    "the signal to die out, nor do we want it to explode and saturate. For\n",
+    "the signal to flow properly, the authors argue that we need the\n",
+    "variance of the outputs of each layer to be equal to the variance of\n",
+    "its inputs, and we also need the gradients to have equal variance\n",
+    "before and after flowing through a layer in the reverse direction.\n",
     "\n",
     "\n",
-    "One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients\n",
-    "problems were in part due to a poor choice of activation function. Until then most people had assumed\n",
-    "that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they\n",
-    "must be an excellent choice. But it turns out that other activation functions behave much better in deep\n",
-    "neural networks, in particular the ReLU activation function, mostly because it does not saturate for\n",
-    "positive values (and also because it is quite fast to compute).\n",
+    "\n",
+    "One of the insights in the 2010 paper by Glorot and Bengio was that\n",
+    "the vanishing/exploding gradients problems were in part due to a poor\n",
+    "choice of activation function. Until then most people had assumed that\n",
+    "if Nature had chosen to use roughly sigmoid activation functions in\n",
+    "biological neurons, they must be an excellent choice. But it turns out\n",
+    "that other activation functions behave much better in deep neural\n",
+    "networks, in particular the ReLU activation function, mostly because\n",
+    "it does not saturate for positive values (and also because it is quite\n",
+    "fast to compute).\n",
     "\n",
     "\n",
     "## The RELU function family\n",
     "\n",
     "The ReLU activation function suffers from a problem known as the dying\n",
-    "ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.\n",
+    "ReLUs: during training, some neurons effectively die, meaning they\n",
+    "stop outputting anything other than 0.\n",
     "\n",
-    "In some cases, you may find that half of your network’s neurons are dead, especially if you used a large\n",
-    "learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s\n",
-    "inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life\n",
-    "since the gradient of the ReLU function is 0 when its input is negative.\n",
+    "In some cases, you may find that half of your network’s neurons are\n",
+    "dead, especially if you used a large learning rate. During training,\n",
+    "if a neuron’s weights get updated such that the weighted sum of the\n",
+    "neuron’s inputs is negative, it will start outputting 0. When this\n",
+    "happen, the neuron is unlikely to come back to life since the gradient\n",
+    "of the ReLU function is 0 when its input is negative.\n",
     "\n",
-    "To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function"
+    "To solve this problem, nowadays practitioners use a  variant of the ReLU\n",
+    "function, such as the leaky ReLU discussed above or the so-called\n",
+    "exponential linear unit (ELU) function"
    ]
   },
   {
@@ -3881,12 +3177,20 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,\n",
-    "in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than $\\tanh$ which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you\n",
-    "may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default $\\alpha$ of\n",
-    "$0.01$ for the leaky ReLU, and $1$ for ELU. If you have spare time and computing power, you can use\n",
-    "cross-validation or bootstrap to evaluate other activation functions.\n",
-    "huge training set.\n",
+    "## Which activation function should we use?\n",
+    "\n",
+    "In general it seems that the ELU activation function is better than\n",
+    "the leaky ReLU function (and its variants), which is better than\n",
+    "ReLU. ReLU performs better than $\\tanh$ which in turn performs better\n",
+    "than the logistic function. \n",
+    "\n",
+    "If runtime\n",
+    "performance is an issue, then you may opt for the  leaky ReLU function  over the \n",
+    "ELU function If you don’t\n",
+    "want to tweak yet another hyperparameter, you may just use the default\n",
+    "$\\alpha$ of $0.01$ for the leaky ReLU, and $1$ for ELU. If you have\n",
+    "spare time and computing power, you can use cross-validation or\n",
+    "bootstrap to evaluate other activation functions.\n",
     "\n",
     "\n",
     "\n",
@@ -3911,8 +3215,9 @@
     "* Make sure you are not overfitting.\n",
     "\n",
     "If the validation and test sets are drawn from the same distributions,\n",
-    "then good performance on the validation set should lead to similarly\n",
+    "then a good performance on the validation set should lead to similarly\n",
     "good performance on the test set. \n",
+    "\n",
     "However, sometimes\n",
     "the training data and test data differ in subtle ways because, for\n",
     "example, they are collected using slightly different methods, or\n",
@@ -3954,25 +3259,7 @@
    ]
   }
  ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.7.0"
-  }
- },
+ "metadata": {},
  "nbformat": 4,
  "nbformat_minor": 2
 }
diff --git a/doc/pub/NeuralNet/ipynb/ipynb-NeuralNet-src.tar.gz b/doc/pub/NeuralNet/ipynb/ipynb-NeuralNet-src.tar.gz
index e55a4b7de38deac349daa4de9100881792fb8b90..8fa98fbb32c06f645f84bae58ce0819cff00ae06 100644
GIT binary patch
literal 210
zcmV;@04@I?iwFRoyTw}o1MSbjY63A3#&OR+#XP|hT}{tIJ@lZI;tNDJx@otGCRuRb
zzPk$bWN#wU{|~~KkYSklST0^y%Oaoh=x&r!dNz|GU-bN(YI=E}O{PjFWvTl!*QFLU
zT*{HC`;BzfrfgzAO8selvi@dWU$)!*(dV*?{N|I`)Go=}dzb6r#fNOSA$-Z&w$U_e
zn+td7{8p-9)9@?N9$Oc!7~hGhDwkdGrWqe%*hll7it(fO0RR9100000000000D%AE
M2OAM+8vrN(01~_HHfL
zu30xbJKk^pY4JpP&gNNlzRxb}%nf>xZ*`tWu(}m
zU5fhEi{@>=NnXukl{0i{|c47WitN7$0R9rB$B9d*qi?Ww;MOQf{3WO
zglp)5?sz-G7c}E^07q{TXJ@S+1ljD{ZGZ(bAU31*v2N;x@XwkoKr(kJF+w)($ZorB
zPQ$0&TUoM>E{ONQ`6lA#hT7A3i&=9F!u7BKb>k`(GmYWgEScho#WFrs=XS#wOX63f
zE2#*W{?D|~pD0w|txUad^L0^N`a|76VxL4KM_>dj@9lC3C9A{I8aD2}UVS7q=EJ>`
zhMCzTA7X|}RJAT`4FSlxIDNpd^4j%SPaX26aHWA0q}$5V{ZzTMchb8d$K
zL2uwUNq#CRV2S}!B(I18#PLS6YI5iY=hrKMO
zt`o{ZNc$f(npvC&kPXZep&g()ouDAs2Kr7M{R_PVB;SzpSRJ4~g0Bpa>g4DkrQ?Sv)IHta=*8YeuivTsLeb6zmI~f9rTI{^xA(L+2G~
z4QZ_^x(eSSzE%*!GS@>>v8X5KPdRk?3YNLNb*t7&qhFns#(I4fSS4Ku#aWX8xKKKL
zgG)~OX8Go5z4vjtqx#QhA1UW96}i*#;a|}ppk(P5{0-_OQWpnNY4E{{vPK~R_OwvnX=Bxhj*<)P$kj^xbnpe=$TWe#*
z=E!!QwTTVTM5*{b#`yu+PQ-oI+MBeGs{E+a2hs=u;=nzc*-pY=nX8Jx$;ly#;+`73
z3-`&YA=HdKR*QbRm>2S|gvvf?fcrUw)cdfEc`j)i6;i(PY8y_Rwk$v#7`w^xFV>%U
z7Y4=h`ml;5IF_U4Mcm-vgs122ADwSfODe3oR%!PcCJ@Ldexp*KuA%Hu)gJ2>eBpq8
z3gTF4i&;8Jc0X4FF*ip4h(KvW?`F+Om>v@ZRQot>-4JsKeLo)W|Ytn+T&z%;wBug`1eZwzF>;U-)c~xxL%D
zXmse3TXdJ&#ql2!(x^~^LcwmLrG_z{ffnnI?@ry?PLAA<)b!&5+G2zZOg^QS6pycI
zt4uD@(#N95);2@=RdNJ~g7Kd!9LyJYzS-%D5uUA867Dys31!>ZHRq?k5H!u)y@aYOrG4*C
z;V%r8j#=qmB|i}W(N1SLdu?9qt9WeLuDu812Ofkm<^5&b%y94oBqP%-leKZG-bk+*
zbQFtK{GbSb*{_mba^h))1WBU4v=+~mWQx+fpTStN3o;?qD
zD8H^5BL+Bd(ir1=tK*$3^Y>VLDc#JVN!5uZQ<}8v{hAI1JbVsB~8cyast{7pd8Pi-_0B$7@wwcDI|BU@u2;d87;-yU9r4u^1svD-T-61`54*
zVP2|xMOY<|#CJ_BKW5KJ{Cwv!a1+^<;}f|Q^oJMjr>8siLQNm)iKF`T
z67E00UoW?i)tGFpFjb?rJ|IuJmnZ3ymnn{%ox1ctHPyOaUJ_mO@h{a|cI@4eza2g_
zTh@ejzK?({Av5+jSHF=OUFxe6L^=>+&czKX5O8E*H;M9`DvF26iF5~G1f5d7xu%_W(
zmp$dygL8&mD-jrTA95Xhotfh)_N4K*LfxyXre!Vxq4_Bhj6qzlR&tGXI$UY!(JEX4
zi3B7K48-tS-?EN@b2WRz%R_``lg8qRhD75&&WcgEx0}lO=PUgtFV|bYGg}fgZ#B)Y
zZKVS4!w{@n)qWNZ!qK6@C5a7)vs2H5xO()MmX?zR)qNRyAU=(i*_0AQGhTxCqKteC
zf?V$`y7|gSjrMI`e^%LrMD^_>S5GoaPbMNiPLjaUtMxE^PB?r7jzc`Zde}jZ<=plx
z33{|fr%%
zoXncp3`AQ}DKDHngUrjS`5Y!aD!h!A3S>Wfa)3AdXTl^Xo+}vQMghx#4{J
z&k0eoH70*FQ}z2cZs1Y~mC=1=b#Ao}zskL7FJ(6*h;vZOR0LZsT_vb3RWv0SG(@rf
zs;@TY+1w`Tavut+2ih|XTj+Or)eE-J-tCu~YMY;<&+FWNx}35d(JSg2tiKOuQ<9uhc~NQ_FPzH++TIf!yG&kB=z
z7O~q`$m7_tt1PvjB!<1DSNnNtEL5)v9u;HpFxF|;)3ma<1f>YTA@oAyal5hi_a#H6
z<46moQ2!hdFNT?CEn>-gNZxk;>_TSc=)uX#H5>1r%W^|mZFq2Gr1aYTzy_FLuzk9-
z?pX#fs=k(BheXC=NYR+9o$AatQx6?&T`J&b!XNfHtCmRRw
zvcXqBPs%{2%KGKCV*#K?Ivy7Z0J*ZYdmMRd5iNO!c1U_h871l?KJe6h0~
z_YdcQmeMp*;iMxLx+7wOh%(QN<~~JxuNeeK%|A`iiG3=e)v7Q1p0B`K!W-fE0VD`F
zaZ`pC
z>}I_%fgxRK>2M!35PF{FBnMjxD>{WhUve@Q#SgrX(n4ihIFLAS(N#ME4=j%b!Z8Hv
z=c0>y|ERm8cDN*-pdQtTyT6xYZ6ZwubCD>e2=1cnuEoaP0`M$9@W}U{%;}w?1Oi_`
z_?tqm^bz>zD;7sQ{wRGHbdB)|c_oo#XDlx2i}XA&RRojb(~*{|aVYCGY6gMsw)HX|
z2bl%ZiG6{#z1J>8;67x@m(9uP-~=I7l>pEV@yClSU_zz!x(5F6)Ee70c@L}8kMA>#}~
zEnSO0K4#8@(Xn7w+J71)Z)BsT#V{o|6r3om%oxa$(RYSXg6c*^cDv9Lnk&PLzC?Cl
z&YZ%F{v3oV^2~xNYIayN!f;qN!gus8!XoPf%16e5-}z=*CFt9^>Ya7gEDFP*d~o3V
z^~BQ`wHA#*;&8N*2Ol52lf-UL-M5vaEgrAT*rI{4Sx%JfN1dBwP+dU+x+aXa8^H()
zdCx_z4YgUAP_dqj%~%EEyhI9}U|~+-$Eh=6=(|jfwZ^hoGa=}-EMyHAff0_zBapK|
z*BC@0NkBKvh9_gT(hP~O=|v#hL2$#Gq3HvmR2lSi2)d-^gkcxHWE~t80hFguSeqIT
zg2JFoUwBX$ZiUFa+}>0y`zDf%f@IN6{jGfv4WgkQyXK5X-LI-IBLdgW*PJK6xgjT=
zxiJ>dZ`|aF!ZDZt+u4R(UX0Swj1Ltcb2g;xv8;b8!2p$9XIpb5%-gH5{$gnIE|iG|
zF@>f#SZbs*K2lX*PN*``B(5+9_0vxryBlHnsWYl8KG|)J&-3cpisJ>R%^n*&t7~lxwJOfB52DWYu*bUt{1?@`7k1yEyKCsox^HWx;3WEA7qYpotZr;0iyS{q=B{JfRNG-ss%X~iW0l$?|is?0h4%NPzOx-;|31t$*I^WR*nFA{z=xw&nO0U-PL
zH*>27BVGJ^GhrY4I0AO=StoicwfEs;_B;2xJE?;%z_Xnpj=jLv>)f})8|Z;dp&ZR_
zI>Z9I-PgA3nrYbA!vW92Kc_q*g0}HS#QmGCx|TlSD6jS6Jj#|Z>!7!fTHhT2u4CV`
z`9pUubu^eYJrV*&_v{~N%)d(zCJPN~WT1L7MpiyGo%SD5B)BY|H4uMhun-F$Fm#h2}P0o;Vh;
zvjPC41k)t&aUD-AcB~hN3B1)5u>vG}q!@*{v>NOl=?7eY7fuyO_xVA<8;I24Y`(!T
zl3CV5L;~2gKWb7v`S2hkb8tZKdMCwsmLa8c?N6YdM8u%n6!%pY$$hFGU+(0=AP^SJ
zA{~Y78ZyO}yt@sjfdBd$;rJw+#U-R7*3emNh(%*ch6lbaA)p~BfvH{}52|hiXs_Xh4DqfPSDxfFm1*m=LTng|Kfn*1m&UCI1t-&ufQGIUY?{L=NRT6NiW$xiGoSo!X`nFCa$F+Frt3VnGjl>etcnJcZ_Yi_E|a7&|w
zt*sQ|OuJd}+14dZ#bWZard(RMdX>zgQSwM~V@^R;
zFuJogV-5rOe)-6{-I3wC#FL1^R`*3{qIDW+Fs0XiHyo&0X__~Fp%>H_{4GOSWg+7>
zWf2&61Q+LTFeUC-XZrJ2X9A+hivm4hGO0J25P(K5hAuPqvlUH)I4CDb6h~bJ`Jsxs
zDpl^EPdmt4RRWe`$N+ism*uZ(JhTAP3r`rn^^lm4gNKg;^!sk=$L{Tu1$j)TR`5$K
zyDHN+J
zv^wk$h6*Z{K%V*8-EiT{@`4Pu4M(_UXvk-M8Q%3!qZF}H1`@aDIlk_~E;ZCm>#?AE
z2G4veK551Cp!rKQqbCFO&MjUX=f*9>7$8h4wQFTM!%QkgOX*D1>1-%`q3A)n#ty~M
zeLz|bznFhYT74`v|5xMzWha5XhZH5NJ$jNqrQ_y~ZfBEIP;;*zTe{$Q2
zicw120@ZMP(h&3_*I&DLN*(b%+qUy8T^3)if-3&#rxd!tP8_#g7mTg2lkE7_T!3En
zWxH1-hR}?&Q9h5g=W)x!p-+4YQ<(G0pnLLa7SZ`Jh(FJR*3o^4KW~cmUhKxN>xd7r
z$$d_^E53kW==cMt)Q=;p1;v&2O}=n^)jMiV{i`!aG4rozB2N`izPILZ4{!V}mr5vR
z;ja9nBT4kPe-Ma*Bo4NR7jGHDIxw+vMW;HKio3hUeZXT$i
zt2P_VWJO{BDX25;z{XAz>K%x+hb6|gEotZr>G>#cA)1CF$(iaGG_zBo1t5N7;Hko#
zMkfXRPA%Gy%+xANE%+=Y@SCDIAr72EJQu@J02PphnPemT3+)QYjVIM1hgn+oq5O%4
z-`Cr_+NA8k;qcu4RYo||b@8XWqnbfy96y6$cTvDowzS>yW<*<@jmIxmb$1|1uTK0$
zc6I>WpicK?YCWwhb6|VB0}vyA9I8Z9ZVz*L(?i_X%vB*-FNGL*-Ei1_&%OHHyvIjq
zDQqtwu&iV2(LKUGuTy20>}iAk5rWo;JPqq~l}F+B#y0>?(uSJ)H8cb|HbG<^{)>+O
z8>FQhp72t-O9!pJntcATv(>`!cU-ac-4enz_(wMO*#zqZRF2rIfh5T9Wx}@h1^JRd
zJ;@;q6WnyMRpZt(!@U4erQu;J*&p_2lfl`zlbpZ&Y#KxK&|r3oOVps;1307%V9d+J
z{%PC+lObO0NvMkGP1WfvNZ8E8%)}1H)^Pm%a7=P$_7<*|#H`%h%xNNWptOKi(pEdF
zU!K7Lzl3Vn;Gy94V^Gp{sWn{-k9#F2%~r*sbT(=auerp8$z7cy7l<n>F-47gXUF_yRN4PO-dwSWvPXufsWC|{Ny1JOrN|gHq44MPQ6rm5IbYksF
zBO{fbY#n;bO)mCSQBM{7_KiTz>4(-zaMlh~oqMB>+QERciyjx>pMIg<%{q%N?OPu`0T
zt!WFQ;^iOn9v)~|hoKp_uL0R!q&BlBQnFAZ@gF*JkYsXK%&P{*dw2vBtk3TvJ?k96QIpOEfaclmbD{dWP*~~XO{c{%}PqxTU_aE}@
zvvk4vZlCr|ahd@87JJt`0Nq=yPBH)wbz4ir=%i^MO<|um9ZLE-Te6h_ZDW0=ES?QP
zdI*b?7-Px+7M_(dLf(R84J`(Dm6=l5>+kB
zN(ki#DcV(C*+C*fg3e31c{W#@cM0%sdMo(*wMEowh5}P7ejReUt-EeSVjzlPHaovu
zZ)HWpsgfT-WtG7+_~YnL3iUM7N-XGv!u0K1v$D{zwfKx*rgryQNV+Odop|i;o3^O2
zYco&E992VQn`m_;z=&wq)>hUYj-1t7W|Y(L>6F7&D5d$DgU{&?ppS^xZj6HDN|KZE
zW@{{!i-2v7gQgv@2GhOZE<5#Ep_1|jhfwrw|MLUvbSN3kdL(P%0UjJ~9q|L!&FjS>
zWymMpFUn3rI#IFMTF5*nQ$=W*aiAs}u=su0y#z~{LnGaUB
z|64SzE(wQ!S`_0w-ieEkim?<@KL-Lnq4x(bA@CB~c4Bs%bWcUhJY7D4>ID=M9y7X@3!iu@-8M9Vv7*CL~9A*Y{+&4ec3+UsMzp+Wx?JD@plWrruv
z2a^$64nCtRC|?m-2k|+Dw~5s@5J4fz1l@JJ;ue>|a-jL)7IJ&iP0P?g9E5=J1`spl
zKaTi>P*jUkc?P!7?}{`umwcJQmM5JY#MqBjVQqV+SsX=3a7*w2lpj2+H-(!TZ>r-G
zJc>=de_`aYcd;|yy;~~I21o>$!s>US4x*cs$Du+j>_vTe?p4FqQeTfrR3XPD#2{xn
zU!K+W)XM&pzIzm!%EG}R!Tljk0U~Tl-aXOp&-#xY!w-M%oyp6p$ERHrp89CjCsk(fw$
z0zZea+IlZ4>jRmi6#D4>_&;o@o65vmeP!;iQlm>i@79-L#qg8Lwp4Oi0k_AEQiWVr
zw?~$N1JB_ZWhv|WbgE+*Fc--3v+TMf^**B@+!`V@2P+Nv%C=p3+S
z%)hoYMXqBACcE0{*y!(^!A}zxvW2$QE!^U8t0r1IwNjgC)2Q9C09V5f5daxZc&Ot-
z{Dky+k5X@MwM!zNq0vQv#-e?y$d$S_cAU1p8a#md;(o2V`rHX6NppKz+nevOmV`b1
zP@{xT9(sVmo`jNY6s;F-#WLsCtZJfhxJz*WR;oE6c|b0+`R)YOi@b&S)3B47vW#-)
z+JoyuZ^=83fVapQV2)Y9oiZ1frZn1J0Qq7fG(bI{D)cvZ%6R>>mcu|S+6bUFbARnv
zeY40rrTor!9ThDtBO=aesTY=eA6Q8c$tKrE>g7>ezvyjXegJP0n8M>Grzp#$CcBlH
zh!tTw1_9skS=k1G0)YZYh|W)}_e*;G+rm#>>&gu96Fk@hYf&CQ_ypbeeQ_DTRQa8X
zBHK!!cUNgHuSR$XVXP$|gwe#H89lNVHvFv~d0hnN|9-VECb}3}1!0~HK)lpUmEvJ)
z@(rGY$TK#r_XnKjv=SpnS`$YphLOI3G5Cy>ko*)kyuJftl|eyjNkgtC@4PB*{+KtPB%%K~-NCaGmaZ&SA(?0-1PTq%j;0^M3e=OaX{eka
z7$?74Xscaq+Uv*WTE`Ys=6`9^mFTotM
zweyJtn9T;ezIhK`lIt2+O<*ESelSJPoQeiY+~GWabrrd71jeCzP^ln~eASNRIiLAZ
z*VoogFA3?{=}`Z`KQ!!*P8YmE3G4L;2j>nH$|+mdjEM}J>C|t^b4CF_zZ$mS9U@WW
z8#?G+$3;5r&v$N)rUl*Q0Ar7NYMcrWIoQl~4Iaw&3@&23;%-P0V6sKs#!u}B`&X~x
z3nApJ0g9XLwb}$1R&9}cQVn*QP(^d}kIL~Ig(q+2+Mk9d0&l0V&xHde*=){}Ea$wV
zIBsa?-#Eg&G#fIqD)C@DI0kpPLNCboHxn7==G6a!Hx4J~{|CA`+5g}8S)*ka_aEy1
z?(GvLYU$6AQt?S}hMAwDSSz)xwQLgpEnN$;mfCYUnMgOWVGH4h&erH;?1%$La@4SLHvB)b}yda8imu
zMn9w*sYs_*{%*=q3m8yxRrQYgL95s0jkrx@!xi7Dewms7R*~bso3@G@
zfTOAJ8N0oY@J&Bwv4MsuR^PIn`Apn5V7ocOVmSvw3h>z1mPQXXIw&VL(zCtdQb7Jx
z#T1}Q`=`x?=M8skTqBUts)<5eo|koKnuA^PCSU#lbd+hf2bgbDV$pNV11-@Tj|g0D
z_5)H3pl85u@d7ZA^_Osl5aH|wt8*eNB`~-ae@%jzPL9D24VWv!5+ejdK4bTQ_S-@3
z0Vaf%o7&4TYxyBNRX^x~z_X^K%*7%?G)rs}1D*JoRM(e8QCVMzdh+UMk7p+rEN>8=
z-&o?*N_`=ctUAk~&Gd8N;eC*W8PZdo0%|iz_X7Z_uy7HHIH*~)6?>;0AKqT`(PSyV
zdu{9&D9wLHh4!o4-*_(r({J`WcxQJ-@*tkrwC5l5EzoUb7lP_E7W--z&ZwDMxN7fb
zP_@*E+c+uyIo993{n{e*DQ^iB+00^v3?t?&
zq#{p$gjh@6HzLy0c7XLmf7TVhEZ;8<3!c^Bxij%n?1%s2sKj-5pQT&-yg?Q>v!)xVMb>Bi3P86Xs>rCkSzC|(y}>_
zUxW_i-057or6FpE3%_IX#C~*@TtoI&wx?v`*_gsLC>E8{bEC5cr^DikFEX^mkJ%D-
zTI2Y%gx)LNR2Gy8n(~yhgnI(&I9hEVzFVF;p>-yA8DWH07;VIsiBs`8;QMZv>*3NL
z!q#_`T%2dWTI09HG-R;2U^Pg%Pp$sJ82#QWdL00Dn0M&5j9g$cQofgc*KCtVuFz66
zD2^m;$?j$O`$7OTG)V{yBjY9f{gq*W(ah2-SfumIV5(C}8U-qH4;2*%2C>^w(O3FC
z0Gn&y<*w+Gjc2erB+l~N#y^4({wxCeV*ou*RLLGq1DYFLH2H9FUL1i`lRUaJir;Fa
z7%B5yR{R9T+zlm8A%ubcMPX3&s2Y2B=^AiOWrsX61aE43ED5BruZQ9c{|i-aa14KT
z?Emj4W3CI$OEP3Wq~RoBJC`|#{?`XV2^Wf52?@{UAcmvY4&r7o@E{idSDfEC>3ke_
zoilQ$icIQ*p&TL3Jh^Lv{lc<Q-aSgQTkh1URF6&5T+5m43UuP!0XCWNAW9F{6jIowqe0hr@iWFqTt_d
zS3y;;;M;u}A{)xSIrOLRMaQ8l4LhgJjL)16zf!b=@LV^Wwv)SVlO|eT9~H_HJQF!e
zlHrx<1@ao5(KV?S=tBG)Y)~95C<}PW=~H1Nkn5|2IW_C%JUPQyqLcw@ll-IyB0qY3
z!H4bqyY6SdwSz41%)*hBn|zsh!Y#mo2g;F);kaOI%*J%=i5F}iwZREtK&P75`vW&;
z&h!BjWE3B9`iLm{;}7<0((fNK)VDk@;+ya7$Ww~Z+H=$ieyNJm+Uer+S-Lp*!Y+59
z>#1`t)hH0CGv@dP&LZ~dk#}qQ#O~r`^zSmntG}iNS4PhPNVruVkoi-o0SfZ^ADY$jJP|dEyvsMPf2r{^(MN24j3hGv-~almxsoNEl$owU9WZD9=aqGi
z-ci21q(K(h)csrd6{9=&L~e^p(KJ$35R9RHrSx$eg2+Ks@3sAX|C20WP}Ek{uY@sk
zBMbRX%&cVe_=oq$;!C{mwIV8=YM?H%`CWX0wFfpis+GF@bk3KMKvR@h{R?2TKl9d*
z`UBlk68-!7k)0+Xa6)+L?arfpVhzOWU66da3>be-0#shmkWDJ>RS3}3D}^@2^hSr-
zU0cnh`}u2FK!3y8p!7yF8VmYVgJ*_JQg%<~sQ99ndwMBU
zxNRx$UlQfw{=X#3!p*_;Uk&d>d&gmm69t$vAZgYlEn=ZkDHPOxHZvH3YQ|aRceRZ2
zZ?Mt{7spN9UjDlBI9855rj9V<(+TKVKfZA7j@#QwXcn-z^_hyTzK<|#0
z7)A
zG=Vz$M`(K9sAMEttTYC>A!hpI--ctUXZB-iBHf9=o#UplV<|+jLB)!*Bd}~{LZU5@
zSw-;IU?jLNRl--0xi0`M!`aUO?IgH116k1D&XmJV33n{db65
zJWiXONhZWFat0d`wUJ@e0fsd%-be4vwV&>4`|@o1Ql2zBOwOek~XZC
z3dX&uhTzkUI47j|%u$UatcDQKh5goq>`;t|fsjOggYb`fEry^d?eIi$vqVV>XAOe}
znZ;S84tv`zvo35REv>~!LXFynoKNqg6*CvtKPyPyA(p@%tf5E4OUE9Fu
z&?rrGjx6fz{{;wo31Yi3kbq*UFf(jjDJ9X;BtTSwbF*1T-{ZoP{Nj$1LHK5s%OAH#>A*7Z=aHff)%oBg>{LxDjZIi4tRODJNHvNIl5cj^xF}
z6;xcQrZ$>sR{F!P{s$yw4dTXwpnKeP1YRsmKSWV+cKDl*9
z!4$+BF*HczXly6M3R_W-^3>U=BCjDo#IG_iC27ec;_&S@NZ!b>(G*(-e>~&gBZlE6Kv&gDNq=gcj&1Z_6XJOOYnojs^a|YI(o~$hS
zkEFgOnN62v!db}VYIwfg599u1i!a91ePOGu%^5(H79aM0;9aNkl$+zB4uM$FFbKjf
z$`M)U=lZTY4bT&i!&n09dQ6T0u4@a}QTFSouY_r2m+XK48bup?zs0M*XwNsr`$+{yMC!fNF6}cUokE1xCinw2As&xP{
z3CTS$`G+)1u13}KiDao$$+jP1%;4;G2nSdxz`5%vCc}l%=n@lGtG}A(3nvlnwj-Gd
zwXQqghouQ#~gXVLLlPWKM4Pm>oT>Hr-qukv}w+|ATEVqcC
z7vlmF?voh`U|A@R?^czCjo;J5Dbtmpg3T(iheHJi?N=f?j<=LvTebi6XS9=_pejCY
zpyX|DI2Ib^R)*>=1P{JTnc6A`6@ZpYtsgXXQdyIQj1L1HJJ+QKX6{}|3ReF
z{m5SE3bBjRvqo>tHRqUm7Kq>6EIp)9S|f<7>r(g3bAbIKZA>AEwHtS~vIkV_El=uA
zIN!)5gJw!?DW_C64;_%O^(soHB`T*`I>}1CBZB|fYzIi;Trcv`xBrMI&`5|OHwkSni?-wGSukp-kGOE1I;sTh#?r-u#{XHIfTz1kA+?A~zsw^b_5))T^yHQ$Mt9bjBwV{74rPPYHak&uujDrIu
z6BO=FsCezuutb%`7e-PTW~4L~ecM<$Id3^T3%|ZL)k_
zQa=WO-!3n&0wq4B(s*OW?oF8s!KftW4RSC?&hAawzHsgmY<)d{ht$Qz*?)gptTru-
z8vsCO;0Xlw2b&@B-fGt60Z~Hqqh0H&!)bvjtEO$!4v*HZGH#
zKIi6jC8^D0Ww&l+I61y~#lJI7L`lb`D-{dLwV5jxMSqqSlX>JE1p~z??v1_8*nSP0ghq>mpFlqv)K(64U8)5*C+9;l40r$1hmrU#4G2$6
zA6~#QRN7x&hs!ggoQxKG!HjBMjjgV3-BWlyVA8>XZznXFHIecms1u)GQK5XfjEpEp;}u+y&%z{hqJ)xf{(ZWX}Yo)hbW(6G$j
zHG`7Fnx*22BEF`>O$&!!S4e#*kiaIrP?7;Y|Hbtae#
znWcxwhL%5xqs4Z?QyV>(kS(Ld1|XwvYz&(zrB&lG3hl%F$=ZQDIzr3h*LZBm*8ZE8
zzpT*iEx*w3jPS~WOPp-BXuUZC_`u!O+e)=bcC7D048UxBt2M&mq)jnz0!|dx#?Dj;
z!NChDW);}xdG^VCnem9@{Kd)D8D>Qjf*Uh}lQGza|5{Apr<6;7bnOJd0jm8SpbR2I
zbH8soQbQ6)6%bGwh6yw2Z{$*6%t>SDvu=u;k0u
zV1HmI4mY$p0jW(b%Ug2z=Z(8XGE^cLW4BM$qp44v78!WZm$qGW_|664X7_E=ytANZ
zNUeG|PSBHJMe?s3bgW3N0B=*7otsr!VVT0C9Z-bvSF);LAn0kKdc*&Eb3Z)=R)-oH
z(Opsu!{rsD#rEyB^@no(ZluJDw;*CMm3__9#4jDE5IDhE%qKK$-{iV1LN6NxT1)LD
z=TsgPI9rwkun_A52OsexEQ$47J@EWW-E7TEun%=FgaWwCPVkKmpn|G}b(Z^Dys8~h
z={2e-Ts&c|$^tyl2cu`8^{w@29leBHrbor!{_pwE^~oOWUw)=vvv{c?pffz%+RhMF9gb<{Ne4k!T^62|RO%?{
zx{%=f$Zw|&aJw5nAyD~UOC-eO9~7uC<$5R+o#gRi-lvV}{g)KUqjepfBf{F_Txdvg
zzR+Ah)u#CGp)Q|?@h++R9BG`xfrX$0fpMJqf{`cK#Xp)77rb8lCPmKp^;z!iLSDDBk)W@_w@(C}xDdqk^g6{@r(w2q
zZgVGhSNTWT*K4nD!aA)mRpIZ7BNtkot9LxU|F}?qiruiT-%%4NG`9u`&5&a
z)o;CPK&;@Gm{9JQhwIwLpMOoe!s88_v2fdg=(Rq+IwXUA;6x{d%bHWICqz0X&L??(
z9eiw2JaV>=l1z9y9a;45OHL9o$|9yZ-GAAzG79rsTjbyx4LwR;SasD;2UZhVn}!!u
z?&Qrk3~8Ew=v+F((>~#1H$xtcr)Jsc+*kX40f@35%Kj_YZFj0pf%3Uvy(XM`*&ne!
zU|?wI7mg+eRf(;np)D`tDSIORc%o7TRP7BC`^C_BNy6;L$tM(
zGGTlSe3fQn&PCn4Uc51Kd88(WoD<>3Ee@pxxYxvJV`ooXEjnFPkZgnAJ~RZmLrm)o!FuD%5hV&;8uK
zT0kWv;A-7WtHp)N_cJ3hI3lpQ|?vp(UJ*2-0AwxF^HcvW)=_E
zPEB-Gkitb>TzWVi{BXyIqf&(7Vm;|Q^i9u)f)!`hlz=Vr)4EL?=`o9@ejmro0nkwX
z$k=7eZ{BGmqmNO&uRV`Z_UgnCV7fx`-_@Hdtp6mKxHKkdxcuCNA`I1$D1YmKr|kw
zzPYtLSuuc}!1nLg%t{}qUMZ&+ToRfhS`zIPCO9WIYg$1HBn_Y?n{?cY*gIEuUewtr
z|3vZg=VZmt%Epc2Rm?HIT6|lgW^4I2;}-UWWxlTO7f=|by4C4*^$8gaiFjfdpFc1Y
zJ62R4Ku^-==hN5DF9-F;De=?htsSB3oJyrH<>#
zFU2?D-(`)$Ss)-nEfc)-pmH7nk!wKXMmp12qO@{{JbG@)owht4Rq);%cSKE8*;>#W
zW}~{lm4&Uwssf)?GPU7?}4T!
zVkX1_XA#2ZR7xJJb9h-IR#Z3%99F`g|$jG_`9C|{!&RMiM-
z#I{n~wF;G=cmGNtsw#xaI>+@@+91l%&co)9Iej|m$(mS*7HTo|vksLvD+COF
z3ymuTBAW^74me!gNa0SYWJC0vJbP9Prg20J=<6o0t)W5Dio9zjXu*fU-eQlrX1*D{
z_IC2Gn$2bINH*eBddG_FH@JpOVR@Bm7K7I-mPxKqdxcT+wzCH|IG*2SKT
z@@UiPVnIO@`1KtSL!aQUZ(Sov!X@#h?kAXYvO@7is}_=QnX{3~A2Oe&&B&M*nkSWi
ziBMTRb0Bg&5zvJfa*u?u-0EGKH@R*Sg>OGM
zEy_P}ci!`J!%OFyTfmPL?6=VTaUW|8TX5dOHy?+84xb@k+>+zOR0i#+Ds@ML(-VDd_v)caL^#X98va2Hr9W
zr|&ykmQ}u6MaT9;9bx6B9S+brIlZ5<jv3Sxrcn#;+N1B23M|^mFq4!iaC3so&OL?TPkceK$
zJ_$H6M9V_kpI=EoDFnW@OBuF2Es3Xe?R_>!l?d0pE}IAnb7yWepN5rSazNEIB)7IH
z5BVvtntnK70Kpm5+d7W(Xen>!>4z>K%F`ErJn1(h+*1NP67Ewfr*B1G*JIM^L%y^S
z&;^b85`V|&+f2*IX5k^8-2j2Y%w1AxUN{Y%kaE&+^)=x|gZDshH~FJ*UO=?~z4
z3^jMZpF2BseM>Ke-!$%-y3vn5y#Lk*=FN7H1WB`tB>=kOU88_P*%(gecq|(_utj)L
znBx$nl)mu|o<+S@D@L`lpcd&A`z5|%oyHDLP2vzXy0l%n)*@bN(T52_fZot8n`u1m
zK=B^}r6PMg%e{w5-vMF8$64LYD
zPxn|AK>>=Cwkn9|*%{HgNo><F`|>mN)#=h!$E5^^^O@EsK!@Ddw+2hmQ+h~4IW<(YHN-tv7@F~>!h_^y2aJ#kfk|M6dY
zfIoJMrz+seh}`(Xh>0U^_Uetlq`D)2-5`6gK|LdL#%z!%piA+uAdNPR`MgOQU1k@g
z@fNBU%yWzAIwb&;b_WeS*2OQ<-1xV!++U~goaHW4B(dE8i4kAb$vmSITn~Z!!-X8O
z^o&2>{pJ4dKX=m=QVMGm;B`iSIGDP)-2L#A+#TT~hOB|i{p5#U;OYpR9hqCb`*8P9
zs>dv;b(l+Ap+S(>JZPZezkAh6oZayvK3=(j6~c}^#0YO0W0HHcz{s<}r-F?U#{Xmcf3a?PhWWc%@WhwpExOx&}
z;yrSUfUBzMHAi?!kZ97omI~o0PG$7LYnue5fyS}wi)&Dnx&k^W*c0CDpQyWNo1*L*
z0xNj;SENm(HEArA3I(u#8@#t3!(e(3`GSue#68Msf1Cgd30zj#b$~h;1&3geDtTTs
zmKMS5)Z8yaaUcc%4}PbTmB_+%urF1PszNU9)XGISl6LKZjO2AY-C%QRLJ$qvDY1Wf}H!Bs)l6`ES7BVSk
znQ}jn7->bM!ZR~Er`OD`;RwOFzuoTdM#3~xIq^=|jmI&M4VXhi22p?|UU5jYCgSr5
ziB<)2sw05tSt+!quk&K1Z%njJM>o7sM_8<^K5@2I+qKqlt_F1<7t>>P?KZ@<&4fzq
zyufC&adw+d1-68Lu8BZ)NB4t~GxvkRV+G64LC+yn{h&SW5!!Q0cP!G=6QbR-oMP}O
zgOjvjvXWK;AKCIJZ4yB2P}0O*^lRU>8@}u8XF4QHkzpdpNph4-xX>C(-GLC#2;#ZI
z``kAB|NnaLKj=>XOi4gx#wT-ckiTJoX*bpyrgq0~AFow^%_VUi`?@4{a(M*1=H(HD
z1rYWO6h0B*zt%^3MI`4{;;6rZh$||lxhz0e0$3sEQQ^i0UDJPq&27O2z>cMBtVuY|
zE7%yz9By?ISdD>9LV{_ljzrfNg7s5YSAK9CT9eWP|xfsjSVGOsxf($P_|Y*+PP
zctk$r)JXxr@P!BVFnAn*CexXwMB5ZqUT>n7vayK6wL3ryrw_Y%mH=}u8|(i)thiNF
z&O~fn7_>ekgtjFyCFk$wbn{t0YQ}8MTjb1SII&s39+C%oV=xWmBK%I7B_w9fXcHGm
z=UY5~NRr9uFG)(9^&KDmU6`N=Ue_D6YC2kw<0!`_n|7LVtQ5^s7=CV802b1|w)EJQ
z19i>ZbDdKaV7UvI{eUP&bFmz-t*ktK^?k&b*?c-wvY=~j^u!w-zIvnINF`)+_#Wa2
zx+F{RI*WPHY|E(=oO+&ervI~!VlB;tam=d>wOovao8f-}>JhpClMzo7lLprn0ya34
z;6W*W-C9d;+&B`x`&S6qi;Qt|_|(G-A7+7>0rs@JF^~Z=vw_?#yOnfHUP(Pp^6R&X
zMXE)G?MD*q4EAD?Qn6U9=T~g=*kvORjGd&dLHl@kAs-
zDVAqNl9`-IWs>WC@+Q4GU&$h^oA&N}rFfQqHZ49)y53d8KtsXKS6a*Thuv8BZ8bK(
zpW~9NpO?A`kn({yZs{2u9Om}m@U**s
z=XQ6PSv4@b`&^T{V7ZV<27)N*H2JD&*RZ!NC90%M-MH%AVAI*XaJa|shuzpXE#o^o
zY<}fRuR+CKi;ChReQ3rl5eV)GtFgZ8`lcQn2^`OUXdEXn(V?$)I1%aC#Z@N-Q#lDv
zmV(3L4yZb3s<~dYS1c&-Fy0?V8t*E9(kyDSw{Bw6*XM;!oe`{Y4}1LU^xZGa1bcy<
z>~kIpMI&5BHdKB!$#u8*IkMj(2Me4*Pw%AA#@_5=^jZWlxjO{(>}>o-#rfxAuK
zVi^L1*_J$GCZEhTbhqQa`ViS!<;)S)S##3KLPO*|gGY(^riSb<0iho2sC7*2a
zD>pp0EhuE#m0~mElo0b~sW&Bm7CtE+Q>>Yvu4&O+pUW#V$T6Pi-?5`}CXT*eMOZ+3
z%(B^;1;xje^)n?v%q)RjR7n)|hkNpCtDcq6(QPGXxiQOaOJG#ysyJ?|SSB(ex{kw2?S>b
zRUw(qRFE@*gYsmB0qGd6Ba9R?+2k=-JeHXxliDN_(ogYXn`ckkTtWa9PuhG&b;{=Y
zDU++fCS+ff<=`2heg4@?~gF3AQkFOKsG4@
z*3Ztm%M;6qX~_Vk222sP3Thxvw=lVN7eMMO1D%S8QV7}erR#v5
zm2tiYUefIw0L-0##ifPC&`OnHS37OHeqZgHPZQ`W(px}cL|r!@IVoe#!;2O9k%^16XEU
zhWt+6n3Hw1RvxF#l9+p9$jIpi}Z%`STK7C&;{N0qrE$sO(fKRbV`RfWe(ot0^s#|sc8!(Xp7vq
ziyRo-G}TZKCk#7Nibfa!?_37HwpzGv*3}NQfOkqze&(ks?N<1bZc+-&NRUD?Eu*m635_RFps--o%60mx9!J7q$YR7H-!+6o+8@puHbs1HN658UR0Cj08WLyJg@72iW$i!ZI|y8b;8
zABK8YKXQWxp2K+IwDN>H*7hZYwc7@O&~$@A|qL
zZ9;-cU$u8NeYUc~D>$_#FUGR3O!^w2V|Q>&8fzCIt6G1{iEmrn(9_T*A`;Qr;1w``
zHC~l4<^D803u9Fo!9?!7#@_*n`5lk|kHF0cfsn?9vyJc`3L}I-Gu#mHaV7^Q&~;2K
zwh*0nTR-tO7XB{G&|MfJDR)Ed@Q@&viN*i`a$C_YtZJLlr&C#mISx8cwMFWJl
zq|ffYv)QSq8^G-Ns{VhN4M8%pu-NmwKsVb}*GrAO7&9kMMUwCPTNcm+2Y6-?o|O
z*tJjLvl0ap%ERxW2fwfGs;S$5UHrPGQ~UuAOZ6~D>xVQ2E41^iv9G}5{^o4`alNZ8
z&^hBwnglQP=udx_EZ+Rqlp*$x2bMb1j)6X#n3K^Kn0+uIs{I3x^zhIiI~$?=apLhuHT
zki4-guYz6O%U|AF?Yw|j$`*XmVO!n1i=1R-&*3GfM(VBim#-4|1qJpt~!k&O}h1nC&jv-KO>)L`R=v!KJm7l&bT-1e})|HuH~
zZo0jFv@aI@q^vW6EIhM+VPk6ujxfrD6Zl3P$GhqyT?{`KI5RUl-}8HZi(mn;{xxwz
zD*b8}h@|;34NQ3<3^xBOiq}dWXvG8XKUyW&b_{f$xcDy-@VhS!PR_XaX5lHn&gFf{
ze38pLvyUzZPEWODptI$t_{7U<af8G2>${t;
z-cgbff{YL}yIW_3F{Gr=5k=?DbpYwMM!}F?I73+5Yv|uDJ_p3$0l*k{;u3LZn
zdiT$--Vv1bs{DRYz~5@`KeTKATvrRYT7TGA-Ip)0Mm0y~l(p}#uttqS
zOsm(jwgJxS)v9U&eRbHkP3K?AX7zspw_H#-_(Ue(T;Y~Wh6JPM+`_U9!|hLXOC{2V
zc)Ia2#$CqbR;>0unELy`b6NJ)L%Xl54lN4K38+?!ybV-UxwNY{{;h5v4}BQ!%7@Ai
zpB8TCu-kTd*us|nwHzsQUeym9OW9W|KidIk5i5Rsdogl)D*}pU;r8pwTZVs?Im>^+
z0zcNz)z-iMu@E9Jw}+~O55&*)t$(`9Xm=ULo#sFsZIe=&lyI&99-P;GOejeqWB5Q&
zH{m32Ml9cqCc^V=X;1yr$zY#2ObaB&owkXi7$VXP1aA^M6jn|~v#ekQXqBh@D^QwD
zBBPkLlmcM|MuiCBlk>6yBMyHAg+xNB=+2Z)&VhYEC=`KFdZOFJ5uhr`O-6uB@By_P
z8RJu)fe8j2Q()Y9vL6$>VWc<(DuO5pWOX2u6Ecc48AWtUQYr<0jKxi9-~^pt*^Fo|8I|U^xW=et3MNsjR^$7&yR@=!?X8ZZv;f`#y0NM^wr4
z`{1Oir#r(Ez?R8P-X;vQAO^M}67&O49k%N?b?{C9q=jE);n&9U5ELwJ
zn~XF9GYHbm-(ngtSP1nAaS8Bf!CoIM+C%T->S?e@VZG!CgBgE&q?a#S3U!tebp8I!
zQsiYzF)rj>M|z^ENBThaeG2Irl9uDY0qMnyNS~Arm9cc@;rz>Z!xDHSuR;H0_e;>v
zu0g*zzamcRUdK2k`Skb;2tZE&fOGHzV8Is%byIG)e}Yj<9emx6V>EzI-U(nZI)NRp
z{qv&74PMKbL5zR1_~?%>v&*=0mJzLSKAC1m#R%d*AG*4EI4(&c35TZy7uTe-3rDpZ~xF~>!OR^~GwqJ7C_VweoUWW7`o!Y1qS7w61
zO}r=nP&Z4LW$2Kw!eCYQE~|k;x$OK@NJZdVy{nq;c-VzThK3lbqJw@+>2Sgv!w{_E
zCAYR?!7r(E1Wm964bgqNvPmiz?9G)aVn|$EE|z~V6i66hVnX`X8&;Ec9**OU0qzO0
z&4%7d6l#U;+U2!Ef{N%4N9(z$P-eh342_oArp;8HX4VH^s0qP0szV5r58+0nVF^#6(|52{Pp%W=lYxhdQlqR;j)Pxml^QNldFFT
zQe#6K%&U4?HW%%QDu?)?1nFv*Mj7I}qOMXh!B*^`sd~Gw%dy)1Z+;STFFU}&s4e}z
ztQ()Sz;r+?n0J|vv3;=0e|l*B#Jx9IH`jlczlEifgLfWVQ@Ia=T96O9&*L6;v`BbZ
z_9>%ra707>KGX+=(n(t;#zD;_t=%N_2(gm64ig!Z96&~W(QY~931kR#p8*Tar`V>N
zXPeaay2*6oIo12ydAU1Pkft#*MMPop_P#HjeZaW%52ti6!)O26Z_3_}HtlxR4U>PB
zA0%pgnDsTj-G1$#V~sDg@6}HCILagCD-@7Ms?ZpoVr*!1OPQ}-FIx&<%_~>U!`sly
zw!_fNhM~VYta3jJWr$mBfCY~8uH03ADmWF)S#&4g
z`q^^752o!tJ=Oijf9UGJ!WPxY-aIeYf;YvMSU_
z8#T-02p_B2`Tg~T$H{Y#`~REig~Um|qiFbQpg6bX*S5|Jq7%;qt4W?Mb1vNBzW~@Z$`Hil#Gqua2
zgpugE!G?{k#
z)bUK)Nm366`HnFMH^9Yl{`#)IIKZ+U-;u=5%OQlc+K=C^R=V|_mE${ap85Lu%jvU|
z*O)m8N+J?Drxy-JK7miZgB={ih+*PHAwz_Q&gsT^>%K0kv@EX|9t{KcYC&R`f2(%G
zGhTccbkw`XG`j6>f!$8MeptXyTJCr&-l=<0>v25I#Pqd#D>ms$Q%WnCoeu^vHHf-d
z+VNH0Xm-xUm+Qs*(?5WFEd!;9MX?hFAtJF7VVkyXQLSJOrEbxwXSZ(Z-Aa(90laY4
z+twG|CjF>hE_PMc6?OGOtZ?!gf5`3C#xM#Z$0HOa*g4H%6)AP6htk1NiV+d)$x*K?
zw6Kpt9GMZo!otIT>`F!~i^k|ZFKoYbd%xs*YSZ(SeA5qn)|0wKY^nMWQ@$9wMGPnyLs=
zPa?_P@n9;eUY_aQU6G
z=K~5hyt9rso5z4Ez4-#h4SFW}9V=lDjt>?8(OeW76ee@?^=E@6Ls9K+T?%&5y`oQYQ0+v#v>t>R4rr
zyr`(UJ5>Qk*O0Q&kzFm#;Bbtv{*gKyi71M2e{uB=)1a~uy;IDo#u60Z{_g5=qVerR;OI}u
zy`KS!Nrb|PjM9p45o7u+yuX5~Up(e8a-`wr$Tf`de;-9GI_j9#Gcnwq`@@<}n9sY^
zgrn4?)jTh<{b}~xM7H!&-oTG?@T^Q*jkT+yTi-bMuGYLY=epvOm&LlS^EvJ;2a28l
z7D%h$?hByFso^W6Ju@ae4^IA$OytvM1qaI;{{Br9`u;?9OB!q3U><8~dQUcg-D~5v
z?Da9Hf2xPr@;ggZ$5CDWgF8)AUp!vBefyMdV6?B5(Peeh@T|WVJ?G^UGTy%`%FSrF
zlh*`07{j&%`;xFpKoO0>Gz`UYtaSYAx|aV8*R3cAV348IA^{OCQwyNp)Z4Du
z6ra?c^Bt8Hl)}NsA+T3apQ);Z(?v|&T2&fse|XyN8d-xtTwsDQ_e#B7i8e1wHBOaS
zKzGG=O{zO9-ia~_IkIjpb(uq}2Ih5N`Op_{$T{M!T@2_n3y{dc=dOWQc;>rIbCasi
zQWG+9K}kHRwIDP>G{j0QqbRiSf)KlZFXB)w8bFe#1JE%8T&iwscEDT00f#uN(ixvG-wjsVor>?P)sK(vOcNy7d3uI?l@plP&}sOxR{
zlo$f~5;ts*o7=35Llgwsx@p!3A*F76e;X<#4-9lubpyJDBq)f(G2Lttf-|(8MSa7m
zqErHMtkKj7;C|KsT9CxO-WFMEASXK3b&+W_s467kdqRi1vMcm)M50|Zi+4T^GWFM-
zt;C4<;MB1;VsWSVU~bI%3ZJZ5gd)ha9dnS{Gr_HExQZuaSWhMrrxoqZT=Ybz8mOHjyk#heIke>}ILAV!45#-4S>y>)FvJdF^g{W#_0SywlHZC1|o{D^$Zkxf2fil;Z~rAy|j9euS;)%=1wc_ji1^Fyki^Zp5Z<
zDVfJ^CVgcw!W8Dc*}wKO9CQD&hbi
zpcv9~4>qViCH!3c;_uO$@Kg4t?+wXM%A1CV+Tqs5LKFwo@nDOCTvtsXbTs3)(`Ww!
zLKMn{lY#XV7cn3(Aa7!73OqatFHB`_XLM*FGcz$bm!XvfDSuj9Z|lYpe)q3X-)Pig
zxHm5M1bJxMBsfJbXk(xSa#Db%Xo)=%>5^0&pI_gZomq;jQ1VHTz<{}2?q%-t&91W5
zc9pHZy2<$Y`Tfl&UkbZ&smo=)x_?|{sVfwG%T{6~bdg#iSGlpNRAzO5SpASRX`SA#
zm2yc*kB6o$`+pZ`>kQf>bq-_Sv|DO-kM!Kv?RMPV{&@fQPrj7JY)`@a$-+Wk9`E;8
znsAXw=}c;!iznUDj?cTgudi&VRF+D)xSv!P_5(K40=`C*KD%AZLM0-5g;1O9)&g_F9EC0XsuVW@dp)8eKRPk%zG)aJ!1*Cqv~{SLgcO1}B_
z?f3WC)XAnC%i&fe^_cqkK&F%Diifq7WHP_5$19=~d7f(J;wdjY0lHIXRe=+9S`>P<
z7O8f6b-zLWt7WO3$SFi>gpSwV)NMC|qD
zlh;hw)_-1GBlVMl@>d8_(g8mT2wIa8#{}wp6w|!
zu+=P&rj!HiSkhwve}INax?5?HMN3`*pUAr5Q-9TW!>~q$*ktUxCw#upr#*aO4mzGz
zmlX1~3?
z?SEgU8Qge0TWhTQ1FA#k#K8O;YYJMEt@$^Il1z4{&zLlYNe%RR!rQK0f2jM;rvRC>
z<)H>TJxwau3N)Vj?xB2W_6>X-gP|zF9y}XZb$t)cL|?0NthZhNa?QC5nS!QbqBz
    ULeC?Ju_`s%(Za*A
zLMiWvym41vO(wQTa{vO-3coIyfbVHYyh7oy92j*#B1Y3W2eE_CIDio5auM&3Io`TN
zyb{WUNNr)_sh}jmvG{&R>0Ts)^EXl%JdZs!T8|{jbPyKx@9!4#
zkpTtfg^o)Jm6ZEEq0(^@mZp{JX1ntMYv7UXhQ2;Dkjr53kg{Z86Yy{lntx{)q+(N>
zaa8#3X>1P72P#RdVcx6>$Oc+5viXJd<~c`$O*ynBL42@z7fBn7fI7+2fP=nl+Iqu9
z6H5Ap+8&T?xLhiN+>0gti@-mwAQBY8R6I}!DUw6eBCmC{nj=mxU_EWo7C7?~n0iC4GOYnqYp>|eW9ZF53sZ*w
zGF&e0}Wmn2^xu^E}>ZQa&=xofrr{V)1Che0qv5TF=I|7RIDdAuLb>_x5zTy@
zs^Ia`UN%VJ)
z|HY|(fw&lgbia@%EFe-j)gl*CM3G9smOQ56XxqSoq=$mt8zL`QBpBpz7;6s2Wd-Tq
zn>ocWHeRJp2M_&mzf=+H9D=EXSgleFrZIVENG}I@E>uf+ntw8l0)Pz5F%y3lOGy^W
zXGjsVbv|3d8;~Ob?PkA?q^K+ga1HOXLGQ9GTi6m?a~+ArwUn<0{Fr{}q)KSONA9$Y
zS=vfL8KHkGbJ2DnOfaH|IU_FYI2qTc*8xhXMwQE`~gc?uE6F
zR`E14ev`a!DSv!$Fe`_%vz;h}%oddpkA>0kqG0|;LKJGrM{>YYw~+Gy^1cSwT$-6Y
zUhFPJ>*f>!Fp5x+?z=s6Faz@j0MnKkq1oWSPt-xFf-^(;p589NVdYCRPL@-Abr@e^
zQ|TBt%EE=Xhew$Rg<(qy;W+3>$7dVl-*;4HOS3{2^?$v6P2=j)2OI42l?o0lTL$C!
zvUp^QA28P9G>&ADz?Xrk%5#8=EMF>yRB|MNHv2gsWgOZwO2tzUSa?7b7XspQGw#j<
zU_Bi7r=u64&#m1+kz!b`f58f)AUIG{n{Wy44;v}W**^tp-~j{sa4-g)hZDMx3_x)S
zdi->YsDGkw$|K(fWBgR>&+^_g66%gPyj}2xGl@$m
zJd~EHbu4j&Y~liBizegNPQ;Uam6<1RTQKQzL-Hn!qLH!cxCaGIsVQjW9M=FkZO?{0
z@49_GJ1(!syK_q*jLVyFPuX6AIC$Yu=m0gJSbu3hH-~V5X8E+th7&PcEMQ1mBw=((
zkKyuRGc51PjdFFMrEimw?m}GS62d`vQ`t{RFsJ5LroA05VP2N8K(hq`!
zM3EMSnZRrx?Els9uW}d9{&tzb)0PFeo|W+=mKE$6p9l|;8`=}>#Q0mDTrYApFD?*O
z2!GPvr$6wR?(ym=BGx)FY2UoUk8(MMlKG@N+Kg-g8{Js$sX4_=oA7x}f3e0;8RS>R
z(b2mWObVRUp}H8ZsR$C7wt2MDx3Fyv0Lr@qQ<}WHmBCegWWq89BtNhbfw$hn}j1ZG2uk9gZn!~_&kG05c^5)GO-^zcDb!Y=Q
z#Vr}f{e_DW=``rkghPZl75Kc($$(v>>JIhrW&xAuF4dV>t)Vmn^8YAkK3@n3H-ED+
z0Jwp|g6f!+1TTa76Cwedq*wLL7xy>+yP14RYj8!kl7M-nbF1ob^TUtXY6Bn9`xSt`
zXFv1+yJlX<`_=b1|Dc}>MD$KenViJAIR)OxSA+rhsvruavar!GHBUfmX9oWKn#KJwRcM$nUa2Ol!>z;+q9c3%dUlxMgu!(_jf%k$Y7u
z8&}!#lBVV2Z7bpf9>Bp|WziiGKo?MOW^5G8V`qnSYtU$*<0H*Qe~h{4KYz+xeq9xg
zXu_QpCQ^Zs85^D0V^Ps!;lp^F&93@`wRMU+@xM?_D)_{cf+pY%)
zbD&pd)o0CE_JQmIVs7{y%P=1OM(N$F1=0&lSou$$zZDDIw6I^t9u9
zYWYj)Id@J+E>kJ_8tg3>f8DBtK9Q(5tp4LuR5eBgAry6YVac`%
zRb{YWXm2|D+;(HIs;HJg^Wq`Lp^omz7Nu$+`dYqTV?aY*gnN+}#V54YymVE|)$w#{
zk>&9%m;lftvm&|$6MxeILw5`})KBA~`9S9+RAUQ&0~D@*?x-uAnt;sti@)EaF245?
zCqoO<_#{Ron`Wqv1Km9-ap`w)0wIVArOc5}diHfS@|^-69<*#Y@R#aUzg>mqKf#QC
z@B;kG3kV3Ds&W)D-|S2-?`!;WDw85}M2s4BC?rwiakzs5jt!Hd)X^YF-2Vp+flQnHgdfQ}|UujHbY{
z$mod!*F+0=y#o#}hsBa}lQ+G4uG-xWIN9;gx@Was`+c51+#|1Na
z-P~;!u*%Sr#ZpLrm0xeHmhj#^i@bh+Y(6aDRkf}6_|4!MweGoU_dTfE?3?}FY>B>E
z(EOACkrF_A0f1@$mH(z4$*SR^VpI^ry?21$J$%AwzHZ%f-wtkY*R{y|qx;n!{j`;Y
zS;?*LUAsl?MVWsa>%vB9S*T0EVR}2$#ZB(&YTu89UK>h(3O3iKU^W6rZQu(jkz?sO
zmX0G#uByFL4#rlw(kh3}>25&2288dL?Y*0WsK4{vth3N<4h2ht_8Y*;!
z;t6It{ETD*S_DaFKWQ`nak=y54O*BZAE*B`FK^uidNd
zhq|*O4k*um`BoL27&=jtv-l92l$2$%G^8k$}=Eix{Szp7!ylQlv3Fe6M^UUuLQRO)MG$J
z38S8@@nk9>W8jj>;xA~zag-5Z@Z>C|KP9_YY3XUnvn2&;8L!79?{Ns!faq6xnMYnu
zNUw-0CPsfsaAB}L1b;XjSP7x4^_K{TiPisq!{N6Mhi3^!Ah=>dYG5*>5&
z#9AehCwcyzSO@%3Ov`*stWiP|I$tC7g(MPiDWpV*0!zi*8I~;-tgeE&GvOEj9`KX$
zAG*9pF|(Hs2RheqI?FE-=o72|y+Du7EEnj-e>r?b99kvI(CXr{ow!BCv7O$gR%8}`
z2ZVO`n6M|Tc^0a(6_*8JfGoo3O|Dpy0&+qSzRwc?2V#rT8I%)75XC)aX-G{FPaBVp
z-MByQ+~9aX*it~DvUL+4o|_O%d-R6iy+2lfO)c~pp}Qjw-cBZMFHS{*tk(w_Xm$`2RDO%(&4lI
zVCC=b-Orl_{eUA|3Rx~uS?1;%zk(GI2m5+(rJ1eY-$8^aD7apZWbJ_EJPr|rU#M|~
zfZ}}w>z1K*B|9(#;i2^3RDB}lgsil(hj7FSvzZdeA%AB7(n~07#7Pu3FW7m;rC<*)
z&KM-Wa<(`m(k0Uz4lj6!oU<{1vtBB6m@*KYF_W`e8;yZJU}Iy%TBljB>%ALqDE})0J*(
zVkPBF2NUgaIAE^?C%3+mA58~S!FhPe_Y%%^I9bA_pV7xWt%;VgxVLS8HxmtaB$^!3
z4Kvhj%ZT8vtJclbr^1;X%6{x2z4(b6+gJ1SQ$AW>%qR@d3E;kyM1X>(OaW}~-As^b
zm(pl#JTRNW;N7wJ1C;Z;B@nF2v;@c`;#F@J5>n9W-J!!ld$2$6F7-15Zd@8T7s{Ah
zwY_UQNQOHn(z@wa#~$r}Q?K1j&wesZK`vq)1;v+xmI|F=YCDGhNv%gh;L)ZL_3rdJS4gzk6
z=CEy6Pw<#GESC(9G2aus#z?X!6!2@i+KterlYFI>&(8aES`@ocXvccz*QnQS9vffq;%+>=jF_9z_J#a5-i`L)GWT8Wx^wGU@TEQY
zg4cD|*&W7x(~*yV-8uZ!rzITE(h}0S?z{|lwEf+K61F1>G?nc#bLqf90{D;lw-o3S
zE&x_Z9}LNg5YjXNVW@~T;$d|Dknk_|(E&#bAhbc_D19DsEkwX-Q%7#VzAZMy&lkjo
zp+Ja_jZivP3{)kcxgdgRSFHQEOqDv=fTUSK@qqbamrvxn&G6TH9|8YTkL=Zhs(
zWs)s*jmnTQq=gzoK;<^{A$`^uhp0LMRVLFwQ-=xBkLAGKN!?CoLwWaiX
zkjFXFHHO=NF)~M*(_Z8wB*uJjx~zct#y{f?$@3Am#;p@YD4PX#stD7B+gxc0UMCC;
zB{BrTXM{Zf3YH8!aG3=A`8BI#tT3f)sqDF3C)g;$phR870l0c=M|G2pbjm#gD(*?EIiV?so13h9AenrD~kPQ-wFQOhkFIcT+A=!l15JSb}Wu
z7eX0-gi*)Aq8*d}V8PR>b}!vMy!*_Fg}lM`d5byKT|~?L#&LY-u2;LYUvW|j#v5Da
zeh6k<*F_*E+}oZ96n7TsC$TSM6b@k-s)^}(h3yzA^PsKd5S_0IAQ^}e+Nej*64!>o
zcfJjyEs|aN*X|!2k|zz`M?$Kwtugp99xp(DxlbG@8qfd$=p~0x)jU2>;>J1_Fo5&;
zW#~~5E;)9oN1~f#+P3-qv+&mW_L2aeDJPQiC}Z;?v$>c1?lO~@bBs~vyPL-qCgmlz
z+l!L@HhuOuPeHnr)HRs@sE6I}cqCl_VR-9-kVk>EK*k}Uqawn9w_jEXcyv$PHYRCe{RzAq({b%YoIAtEAanmWE=2BFs*tA&TsrMJa}9dPJ6%Gsnf)t~Oi^
z5-!p&Gtaf-3Z68u^?b29S{N^@!@?q0b7_9GXp<%h*OS{=vU`82c?vXhe9yP%o7$a*
ztr&#&+vUj4GIlv>?FRM~bWjxN+sM^Gp8a->-tLtX-!h>tKGg*lcy#LYLnD2
z%a=^Pc&cS8wig*~@}?7%*jW
z_%lx8EEY2lzyxfFH0;bHc2UoOZ7)mZ0x!aRljpr+BpogNR~!D=>XaZBsO5aiGk_fp
z?aFyh%i6!puFbihf!V8cw@sHdaFJr>qR4e22qCgOPep&>e3O=nFkZ)L4KUYUr$Ddb
zLciSWr-l);*CI|*y?d#s6m@4!6BlxP06_>K__fE)0VJOwv>eJ%M2Hej^kkk8F~|y7
zE%>J=(?rO{(RVXV03MC3&Z39DEQgT?9@LvX_HKqD$+(CR&1T-i25=krmeX0)!q(vC
zA{Pj+5S@RSl0)o-XEU{nSVS5f^4`=OnVUiqB!ogXsP-~35Q&0Q#*8AD87BM<==`8}
z53r<*k5p6WIMt5=OheLT+@^g~0u(~&!OiBUm~sdMxB%jrhaj>pgNinso_ak)18;z5
zdM)8WK_(1&j2TX0*s3*WYQqT9f&+;s($zc7A)|k4M=&!}5l9IRLpY%4k&02^QV1`C
zXk&>~FY`OgC<@`Vw!}l9)k%mv58Ojg;QcS)d>Lq0;>>HdtaH~LM&KRrBvu~+dOdUI
z?{~88gLn^tjqlcIq@=C@pdNry=ii@Zjz>MaD%2CwHD4D3()L!r*o(YW!$L-*9Hcoc
z;`)DVonvDrP`0dN+qP|YY}>ZevGWEU+qP}nwrxAn|0vR<@agOn^pZ)T344oylm1gSAHd8|J)nJU&!r#kv;
zsuo@+RFY0*uznBtbPq&}7;wzGYtfk+{wwZoxf1`kaV}EuU!*U!`m;VqN17&n0!uCi
zG!!$qW#)yk(nXCPOVqXODhTxtC0L88CVOA3h*0|3+Qb%(HF0NtWy4?TxBC4f;-i5U
z_p6T$z2a>KF;$-I#LQwW^`#3@7?&7VPtlPDNf?O$_=g+UJ69iCIe-TQNQgXP6Wgl)
zEQVvxiSL{ZSUm3zZn>_*qm}J%IG5ADE0t2-?LjC8+U4OZSa_2o{EjOADY&t2f+pfQ
zI-G{jzkW*4TvK#Q1RG0C2-!X=D~w1An!;R>YLddVbXU!@n1GRYt$^LGV}CySgKPIQH(R;S!^<6i56
zjv5zn3ys`}<2-KkRdyP`{R5oWGtm#b`u1oiRkIHsD1zHgY5F9+iTh+J1?v_CD_!}^
zzIT0v-E&voMo0Yx?;%vra@kd_c#4IYz1J8eNbLTxxI?3@Ex;W-^IWT#qYe{9?6&)I
z2Hsf4HJTs-9A?UJMc~ccU3j`#Z|{6M(p!f0t_eQMFvmPApT0d0);yS}ldM#XU?b|V
z%4d@OqBO?}#F}Bb4y$W3Bv1r#stW%yu$yRbR)+RjB
z^f49c;wJDy&6l@+d%XC$!(6o%9c=R{GK2Xpac~0XSf-Om28;?HX@ZjW{niyEj)?Q?Rq)Q;4l
z7e`C%Km>L~Ki2URLH0un0~v5U_wRgA7A4~wbQJrx-K$ZT&Gnxr6N~l#xb3sBbFilF
zV*yhGus7|u{?jclRD^iUp&f#^qxRYle9;Gdfe{SEX$cb`6O+=^eiVpFyIFhN4xg&4
z6jOv&P54gacrrDW?%X@ov9jY-O{m{~a@kgM%tQPgq}1BVKm_>M|EK!a_IO`<<|lJm
zp@FYVXCDtwK!ec2ph$?7yQ*mZ)baj`(rW~~oleK1y78UN#=K2PiGa(v1ddX*btgJ+4FTJGS
znP^}q#WGd+Q2sUf*Dq;*1_NG?sNwk#OQk9353=?%ZEz9jIsV{t)$mD1V)e@FqO}ZA
zozAfk<1k(2`!p9yx&#|bdy{CTCvw5X|9n2!tA&+Dcou2pqS?afaQNuDW)qIdW~lJ#
z+uwBv2gdR$Xvm#GdJNU9t{!tF_Wfl_GDX?=@?Pu&Y~M`|sA(o=XQh8K()hd$^9&ba
zJnyyXWZbAvi|G@L7nsQc-E!
zPl-tcwdlCB=>2YLX^dl{l?`Y?U2|_$K<^V{Uw+YBYC*5oV7fEZF_>t8Xb=Vi{W|Fh
z^H$}b;=rfwo*zKJ*W;nPLo&d^lYotT?l_MHWcMDG3e1>Nsz}bu-QTC<{?`X9^iZtG
z2NdD#3cjjIXw^7-;e#<%=7OY{>Yd?FBJ59qaCrt}`
zj0qJB4#cxIkw`iA!Wl550Ji9#BRx4=b@1@M6EtMoMw
zst-G(p|1>R_VL*7@y3g|f7H^HwOgE!X1#pgk%1;)OHTg;&#e1Xy0$eagxc9ZbSOr~
z&gLKtDP6Bt2SfRmRSp0|FRFC#-PTe4-RLKD1xd!S1j4l!H9gj*DWp%&7V$p=(!{$_
z{hii(Z@{*mA;?f}QF+Pv@Y=6yDk-_5@>vwJoA48?k|>$ME9&i2o4hU4Y?>P-{Uu#v
zj*Bkg)qMDz`LNG&xmATfeO)*2t}?lG#aPK*#$fhE>ky4Q9sanUE_nCen~)0T8H!r3dTeb6ZmuoE*)63o
zv>8m*ZJGc$BRioDU-l%F1&f^cI153;Q78xuK8H|bTirj=-orBoVGlE^+vUH5@4=$1
zye>SWcT%ewYx6LIvS*J3-rrqV*P}ZNGDDlVDa^=H+fgrgHNruAR(s*lKnI<1G
zOIz&y6yoKD@SQ|mFM=$w199y51)IAjL+}EWNq_)sFwWe=SpCnzswa89O1SLe)%N3%4Rc4Vz6DmSgMCDydxrM20`h)hj!i
zDE@9*Z&jJ+Ny@JG*SaSe@467wS1vQ2T(tp|OUt>x5lI)vaDwNZJW)hl3Dimo23W{5
z8z!{hltUjDSjs)*BO?3~BlR8XDD`c64e;Xvigi29=~ZT@&C_Zpd2`ns9-iAsIwG1|
zo93m_+>3$PLx1wGdstT$Yl}GOxMBd}XRIwV{VO@DIG1C&JGO~{W+=}`47aYT+Mf(K
z%B@Ril`}9Zb;h2I0v4?=1Y!2Cg0La8s}${^!^^5%Od3~(=79;tWZLt=;`I$ChU^m)
zQrpXMca#{A96_+$hfdLXJ*LDaSRFtTpNi?)iM`^Kn?L^;l@>q`iIEn=zpa0F(yLo-
zMDfoXH+O{3WFk<4mY?jvM2disl&}KCb3VhB#KgI&p8zN1P@b&|g?#S#cM1&NKR*lu
zr-cvneTUObE7=bUhJ%iM0CO`+J8VncYxT#<3b=5YZd>h03Yuu<_&UhBsk=qYvPQNsr(i5E@KfeqsBGy8Yv
zU^Gss{-wH@?8=i+X8qQY3Dn!bf^?QuDj4H#1xw3nB3voPph--;ao^bnT~M}AY((6
zyX^qVdXn`7l3(&kS0ATdEFaYs9Zg9Zk;A&$w`8ux6B*R>a4IV+brSP@fiFw*b1rO7xJ8RFf|n@|vQ(7|g0uvb()U`6`ad?csPf-!6SL%ghL
zS8zIX^YjWC&+3HRxp<_QnKNHDW$aIAp=VgI=V@nCj@<=rj4U301bD|7R4iuW^Z=A*~>tQUV1H6?8
z8Z{eN5_#Q$eI#V8gPC%5nNXKPz}*N(CUhu&9+ahc2V>C%5S6aTihi}MkYbQ<#Y@#p
ziegMaAH*OKI8X)eg5{v2X*a%QcmQ@R99
zxF$KeCUkF=MU}j-O)m5A$NMl9Aal{6`9>Ru7v6bc;p-djY_Ya%-=%CTfB;(?U^hF&
zEH2t9R%3sLX!xA{kKl*s5GN`+(wr2-35+5nzB0E3Q^2Fc>@YQ
zWw(%OKHbI$jY~f-Xtf}wRQ^mj+6_`e-wf$(oMxA2p81hZlq^(Zf>1VQUW3%bF$1$b
zoM(oJyar#wHx%>`gWI(_pk-CTVT+US3+#i~v=~%`k_Sp=azI9oSS&Le=A8;|z=dkb
z)A1MNGukq63qGRGz(In!#Ua5=FpD3E10AZPW_TqsolWmFj?zKk@ozCgekf$e5F~4~
z$iAi*2%&A?1G_y4!ryDF-eoNWIgbBKh}`v|&XS1KMnQzHFsBak00mbvPJqZ*YuzHW
zV6~!NB2+f%$QK`D&xj?ECH+1#foYCM$4y!shB$=GFi>1L9-adLXWcpWs-5k|ajy~F
zTFZVgh}4_(_F$;9ssEt{1f1r0KvG&rWEYUd$?-Ju$Az^?8VW@SyZIL`z+brOwGX
zGO4)XMY+6UC)wY9H)kb8b2|$zp5M}R#>$y&S9krW8GT9XV~A>LbQE&Px)Gw}OOejH
zqEe+?O)!%Fxl@QZvCC~@R`LF%S*bjY6-zdoq`pMDK~f@zlhGd6+)gd$I*}XYc|mSB
zZu{1p^9fmt6dduN;Q_+NocOyQ9hmKZ!(x*I&2P3AsSeY#QtQ58GX(C;zvPS{LJy=$
zY!f?5Y}49oJn^V)y>@qBjH8&U8lSfw@@N=^`r=Nv$GPyBTV>Q)KihK}-81Kq*x&2eDVSMlKxE8~X0uat7
z5aDJ6VL9ks_JkrkwLpgaW=XzqUR5K$m_S)RX47p;n<0|>dgB>dlQD(K8`nO)k+u40
z_T2yYOXOs8gSW_hCk93Tv74;TheI&}$xJJWwAwlj(t$Dd6b>DUVB|uBT7W)LW3s)e
zwVJU|_?x`ijo8WV&$U@uSdrd*x?_nB`xd
z?iDOF=nUY^1?TvO1SNA#hJa1-s$zNnzMsgA#!8-r&BgcLgL&YNe6oy^3cyE$_D-`}
zn7XK-v>%3ja5Q&^hE9MF^AZWO-6N-l)U@42UeM|tM7UM5IrZ1oA)?A^2cuYW!Et->
z_l^4ouycutA8n?_vbL*#4^HQYJ@Xcz)*0c@De@beVY8svol_
zLfyH?5GgLhj@a7+YQ@ffDdT*dLmX}dKK-*VbZ~+JQsSgLazczb2cTx7xI0@2EyfS8
z)9%HoQ8PvPTR>U>Uu4_I+$$)S&63$4+L
z*0B1u=w7U8>O*&ODgBLr+Io2Cp{_ayrD>`_BpK+NXxxJd9S1Ay7X+R7$5x>ZHK<&(
z;;%uy!ue}kUFX4*2Y@4iPrwFvub;eaPiy(UC%@PLl#X`r>INno&ZNsQ=cGJA8&Qou
z{$w%Q?y-o=-%MtsPP-rgUB@~UNI
zLa@7{<^P7uo(WV?t7(4o@)N2HV*V@2uN2PqjEq2(8$tz>4mdeHej*}bm`&-z2qK0j
z&OrM;Ja=H~)ukw+6u9S{6FXf&vbw6R+T3Hq1X)k|hldhDGEW#Xf^-l4od{9|<~Ua5
z0Lci0JJtEkY?27XSTI1@ox!7-E|_46)fx07aSa{xYJ;vX$e5&diaw<{XH4vue{m&n
z_*TJZ73K0+7l611(yt8}!`wXBr>h~=Je{Dpxrd@JJFE7}ss%$M*=s0)363OGH19mk
z8u?VPg8bxv!+jmd=kcnK&zcfnerB+C^
zIcZ3p)Lx%dIy!y}wX+2$T&ZwJ@G0Dosh3tk5;=dw6d*mVBr{##S(TSD%s_@e(G{}t
z7@6jpH>XIERsgkefnml^@H&Ja^lN`$l+(S>N^a0%&j|mfg4t5Rdbsqa7=s@|TTm56
zwXlh$@x-f;OzR>5!zP{Fi>e1}INIjdy_Hsar??T2zs@rv6G?nMfFZg&OQu638kuey
zwCKcV1JDQC#%NS~C635y&lDd%g!QXi9`#qAAY#sT6CMPd*3y=tN4M*mVP%tVXj*l0
zN7{#QXb-R)lYbIFKj5elGojf0WmAXQ!)#G8`x1RgoEdBh
z33c65$vp7QAw=;Cj0)jcz7DRS7p5f)yhL|0(-})r-qE!A>(1J+TUznY>Oc->K3{_;
zQCJ-i#)wi?YWGdNnD|Gtyz00Bobk&oftM;dSmQ&91P;W^MUXgheONC!V_$XyEK4mf?Ln6S;
zb(+(n)Lv*}dG2>BTprsZjKXx^ce=?7YziNspC6dAPRc&pjJ{
z4B`vBmP_X)thi2!dB`T>F?_p7wift>wSc&lohov7e-NCFOd;HzQlYTKeQklARy0lj
z{0TK}XjAnTo*H6etT*6g`XZq={;!SJy4`oNE7zVn+tz?g@sBdR=9%NiPoh&c2LRYn
z=v8!?hT>kC0E>7DKj5-$UP>M?iai)1kXttCSQ7yhf*RNhq_}
zz+7kQyu)G>qLt}q6m}EXHEGY5PwTIVek`$`TsS>r@eCg@U}f1zr&hq$`9p
z^``Q5$lT_24oJ@Sk?XtkI)0S)E-l#AIm`!(tp~F_*p^+#$+S=vJ{h$OL_K6OYI4A=
zs4hyYegNDL;rCA%(SMJuAe6x9pzNF+|5J$n)5G$=2b>44E+7M4Y+E-E5x)f;D})cN
zzomd4Zo}NTn4?@t74&8G_KrlS37s-!n*;P6Nyc={`ILJ%e_)RSDu@%)=uuP&uDex&7Z3$Uu_us+z2)K!HDwI=j(Jz
zUF<>Y$Q~Dp>uOE&
z;W*jl1j49aDEW>k>`sphjxp?!C!@cg&25kJ^^z;RnwKx2q{Y89YTd>F*I37zUW_p@
zD6PnWH+}}$7zHYPV=kDcMMP>@(6ih`<^@%LEA%mIxV
z`Fw;r@dk*{1XB(+^l;+|{lc-U5nF66hGSHhqrpvq9)dvCWP=<
z)d%
zB1stv4WGD`Rm(~8rD;p711A?iF#ETL{|Rd%C5}|Ya!@|u`e`TCC5cMl+oq4^O@f+e
zY|}z@jqj(H9-Kx2P!vCxrf3I(yu%T-j!#A-eHg;)=_kv
zG|v~gd^MmrLlCqA84&(fncDa7Ta@*5dRI9QaXRN8lE$`oAT8x
zkuu(3o{i0Oha$9_(G8--J4&l`9gms+Rzuw5Kw8w{P&q~DiWwO?fQRh|a-f^piahK4
zg;wO3tHgyHB#@mE2Q25%cCFi$j_~aVsd%Rk<4H0CC4n0Vhk(3j_MFdkwfi^f04~@c
zXajwZ*^Mk35Y;*p#q;N5lUc+n0YJ#>WigAj@Py-+JNc%SuukqT)6r#mfsj{i{xd5o
zrF|zyXd{l=zb;(X4Rsw4N_Vy4t}LLOAxNr5Z60~S)(;0Q;UK+i%UmA8ctw_>!$QFg
z2_)i}<#j(z8EMRK;ZqRL-}jmLdX3_|aW@Wka4kLt2+WoS$XEugiQ83siJkdnvjv=9
z96?~VB)*nZ^fL1I^nQ7y9Pa(usdMu6?Of)k%-tDTZ~1!fdw0j~o>1Sci|sm=_1<;*
zhF(s|)Yse`AR9o2ATg{A!pw-{l3#E>I
z|In%gARFKI&_ntNIsNop*A`TqH@pg~5qmiBZEhXygAEfDec6t61`YQqqaRK@H1N0k
zq5bvMEV$7}cyF*4aib|Z>@ZZVwmi>j!)W$RS9$>m@(#FS6`X^?N{s1;rI*B2I8iS(
zq;=dvRaz6@%z$s`@l&Z6$6fpz=nS=-+Am#zni4gKkuI+?V*Aq93$CR4A6NhKnT4$-
z8BiR~+$%qq?XQj-Z~Gxxlux5AR_qtDA+)E9zuqs!CTN4NHqFHe51BJ5f2SCJ-G}78
zEo^;+?2$c={>VJuiuHDCo(T$`c4(bh&bn2`(WBzb{OT{6XwKCKk$2zdFWrW9%j~3P
z#mKHkVjqrc5c?o#>q2y8B)i4$}D0-kb9K<#X!OyG%W
zJ(G4)qAJksn$n~d_2sOTs=(LuQgOzoJSa?CU%s`njEQ^Jgpbw&lA=hH+YWV`)Ca3fX@
zOyWXNtRRScM_|EsEC^5}(9aaK-N{R&eMgu!<79?or~oxs^IPj^8U^1ns#71ZFd#&F
zdONX5fKu9q>Kdt`_{BfG2Te7)VoviNl#o?|r&5g2
ze0VVBnlKZ!k{5`r&yM-S;
z{{*E!Ys^>brLtomhS3i=d*wMSY6jYC?X^fh4tbQTzk01&qxwVzK?(XD2ok6eqW-Fl
z)%j^5*OOKI$^^4tf`o7B#Xh)AKHsoN7o$P12Wx~#?)ju}nhv#1{opA?*8-R@EA|5q
zz)r2n?vqSR=_X2{l3S3nNCjrgO$|yVFYFdXGH4ZQyFLsgYQRMV=Q>R3>W6(IY?iJ9RzKxgo<$L^uA
z7DDkjx?94t@nfZ&9Erj5Gj`y{P8qokzyX#XBGFR9xF30<5UL?^MrXWwA(F#2(_jFo
zDA)y43~X4-*2oQ8uCu6#K!!bGOTYy48Db<%8653=}FCYP{e
z<*}zI{_VRUpuhT-g95weg2t6@J48~?GYcQIQ?thi5&qcV>x$+Nq++I+4EOk
z`wi+)6^t{r!y_X@k~Biy#k$sc^n1k4+W&Z<&5WS)7y_*{xop&B%jtTRTwt9C8BYmU
zX}9Krn~d2T@wuLpoV{hSmFL7^fWK4R
zfgh|7^lv(6rDTxvZwKWrpWEkJ@R!)ZaKIvEI_|DCH%PDy)vb9*i=EDwY
zZ(V(0PY_vVUpiC%vfKa$nq?V&Ra-)!;suIBYRRlS)V!`vr4K_MFdB_hyF_1tYophBifHXhHttmP9bM5Zpg2c8#02EpFIpx6pzG|e^+|d%0yj2?
zM5&uQ2{UaAJ9jTjKI)EpKlT5L$7c2IDjeQES_yY{!)Fbkom?Uuz-sLoehZiFNiB$f
z?@%%^b_8mYa&ntQ1wc!hQz6!H+k?GuS<~Y3ZhifJeILk3EQxR2iheln>)-`WbRaYI
zmOPi9FQ653eJlU~)ZBBZ)4iM;VPc{s8oMGTHJ@pS_VWRXKYCxM8!CNp&-MxM3cx={
zPI8kixO`3Cma~|SQ3XPmikvrV_H7?Np25-b{q?+@mU@Pb;!NW-?#foFyp6rE{ioFm
zib`~d`hIIr>n75pZtAeD6(2DRes*>Eb#Ha#7uea!Gw!E=X6H3>cU_NVdx?L=jx=vq
zc@@+<_H}qS_}fSP+102Ied;L_L6|j1sU2hAG45CHOk=p5zPEVaGK0RQUWTP!)2*@n
zsiWS<1m#2~U+Vz0KRpMUYWZq~XBNPo>EmVeXV
z0vgI|za;`Om=KgA>NzRq)qV-x8BW|LCIWzm6l@ax|3%qa*#9kiEKE%QTlO@yVmJQ*
zT5}DE!94TgkVV`cea1w(IytfoEuy3SuLd~aq_5(6PV(&%4wd}A)Lqq7&?LlJuuj*4
znf*2Oe0^Ot6zbOK1BW9k{d;FuGzr^u5g%
z>2_;+J7GSlG_(As?3oWg9|6P52TwE@^)K*td){eptC^YZ+v>fmUp8OQPYMo>@Kdz|
zOS8p!on9(uC)@pH4~e>OTxU4~yn`Y{UYotp2YUv$bR0w?ka)`}EcVM!0QEH>yNHbn
z9r_fGwdx5VsI7&!`3r)C1aOew+22~eD7qdyCjPNnU2=>36D+6NaGyXHCZEtAi`bzY)zf(d7uWaSPkkYE
z*|Y0`Nec0#$xOHp|H2F|Bl&U$TULcK+rEzDftU>^hkv}!{#F21FME_x8NXb)RxNF|
zuOvBl8>RusZV%&zgq6mH6%vQv4{}tb%a{h=F*rQ{kfn+;&f{9a$p~L3jICBeq_fe9
z?bD_&Kw3*_^uTG=myY@MQS8iy;e~|x)n5Rm31u$p^LuA)E+FzZa)=+nD9CPMCDTiz
zr0`4jyVl6|OQVoRarF*Mbhmf_t!m6oD$8H~Jq4Q`_n~EY-_=RJ2i3*q6FH*86!net
z0_yNr{qzD-BBPEmA;oy2n7&!(IgC$WKI$*Fag2DR9$$`C|ML>Ph$#X})K{Qtah~AE
zRFW63qayv?1;7-OL}ymL6{q~RgJG+aisz_d&E$3dWw!m1kXs^inpjk>;aYIrH2LL8N<{-w$DFu@jDC1~T2=&>a
zrHe8QKzrcRjzE>QV|;T$v5JGi-k$w{Ba`QE9}jBIGYlUXYV`EY!$
zg-UG1K=2fbz%*+hV%;IPZ5a#;%$VcYkI@oG1QeTt2w6L`N)0wd+1VsQ+D##27vcXg2H+#wJ{oDj`nY)oqVHL@lh
z8f-dLm;=Ta`3Iu6oBz*j+v4>cvd%W<6YxycTDknZ`LAf-8Zpd@82i3QZ@rW=Z6A?g
z!TF_=3jfdiD|b(Gip5y4?-6;Zk>}yML>e1U^J!l5HONwVi_1=RKbvU&s5Q9
zKi<9`z7iS@^A>E`W&n%Wm0cee91+1a)MP7`yXcKVIn{w=q{29b(?l2$IyZ8h*q}k<
z{2da<=)Cg1E(7$rfSJrX81qsYDulV_ek-M~fvTJPHQWf6a4p??ph`GA$6wq+&0Bmj
z=zqboQ@Z{w4T(EV_pTc+^VXB3lD)V?v)JGS6etcwBe?v_PZ={RE=EC2*zx0n_}tL8
zF`|C)d#nQ#J>(+XLO$|AsoA%HM%hWDu|b4DlS#L#zaxFbmh8MWunWm@Q1>|(z%}Ybf5lVi6D%gL1Cx@
ziMDBpPUL0Quqh|N;Usp=t>hpxmG|z0+zBO*Ht!KcQD4lPv&uPPL?(W8>!)WBfq=!<
zIBs=3Jp0XeCPgvv=>4f`YHL@Tzhf6KVzZkS5e>kE&P|sPJaEpSETa$NgBj>h(<7i)
z@;b_|?R&z!j@3w6h=|kupy)TKTuf1dK7%4tpi=L^Lf_lo!fi
zHTwbW-8s4;zbXo@VUxQ?aRxu}!ov`Wu=LHSJ%
zEHuMR_EBQ<6Ce8|B1qzOsxq?67^?UJ^AV$Z>5yB|aou;g$OWq_)M-)0G(hlNmm83`
zv08<5F>@adVWoF>`4&q}5iq9HUYtquJE@##nuv7FI*F<<(FxK-kUkpAo?ke9V5Rlz
ztWO)jmYg&dg{%pY-oJtuCwn6}3>?DTAtnote3F(JR<(%p#kyVr`=iAOg^+!bBZ-uE
z%2)udfVpXQ+K$x+jjU|e9`HSjd*ps?4eMy*Er)d@>$v7FO3L<$15U3r$|IzHL9#g)
zD3wDLa>97D(Sb{9ywmij{I5_%I4}?cugAsu%(&S+3VzqRwN9SG)6WtZtxg4)zMgQb
z76QkxiM!99djCM-o^*MiMN8i%FSV}n1tycpzEa-6m)xqUH=229^5+;*31iEgM2%#6%anM?{lfjrR
z4e{S*5!Oht%jxeRNI+I3-ajlpw_Nyrk*U&ZvZc5wRf8X*6X
zC!HH9#_Sf7+gnaXfTnXQL>p(YjVKlO0Bt^lTgBi3Qs_?X3dI?l(Jn=iqg?U>-!j*u
znAF9n;TqkG*}!!ynuX0zOV2ywfW=Dwe+?=e9RG(wg_G%jOw{B5s@7P(|0D-(wD|e3
zWYO^in7{eBE%`c}V409`l0p70R+*J#Fl6|nA5VvJF=U-mWW0I--I`=#zN?S;`d7R%
zqwan@kJpc#RG9?XfcyX8D&Qq
zs3nXpV8qn`tJ$>{fr?)=j^OMnLok^EK4@*Ka=T@OQ&X3CyBGHM9jQ9?735vDN0@5z
z=4bjB1=NS+zZTz{*c%qWGNXk!nsByOX3iJIQ2dV!JA0-il$c#1^q5Box>`yxu#cDS9s3
zTGj|)*b^p*Ar^xPMif~b>dfuTBAqWYZK~`P;KF1kv6rAiA%bH&r1CE}@S%
z@*-Xy!)zVD7@fRW)e7`#L`mU@637+88`lHKmldLR9UtIpEo6!}WW_gYz`zZWt5-M2
zDk6dz(Mz`!pXByh%leH6N@#wqV3t)t)+nfS!5IZ*n*r0qt8xicHWN56+zj{UDNekk
z>elBJj5jRS<0^=&5nLuit@iRjP(Ki^2-%i$#N)%mT)13wkWMwBOItO~pG_{h>>rdN
zXYwIwOwxFgs6mPON==qEkQO>TbDkT8XT*X(u9+JDei||HXp`XWi*k5S$Ym@ROVrOW
zskezA4c%;@Nv7&AaSQu4%rt)?`H8sQ^mE^fjaiS{E>8W$VTjl6y~`+73l1VwvVf3|
zu!-F5bVZ_Wz4;o-=mf`?^hTmW!g=IFTlHYg>(~5X@(ujUCmBVT$SA){kYWHYtx6gQZjo=t_r)=$Ia0Kv&6!
zkl0T{a(W}Efs_=5EMeJ*meF~j63@=Ydg;ode8)TphOhXqGh|3C@7-^4eOKSk+81p=
zjhm>~o*IJO$rZ|YHB)bcG742G3o6%v%Tte77&|m>TX(mglrM*}0_w#;ixUbP!i3IxWc*51$d7Gqrn>|)S
zk`CtxRdV}jnS9sWD2NKuvi4Gy=7z_KWrXt%X^8rSl<^uT{s(f?pGJctx!!MYR&e7<
zobo}pAfDEIawhSKP$NQNq`~cuEyG|+$K^;0<(+7^Ddn9sOiZ^%nVa7LzeZx`dh6&a
zCas!3++3VwU1!}uYp=Zo9Rd4+_6>6AEnEUR@U_Lv)MX2SOyc<
zt(h%5=wD4u@N2~@R=jnZ^iE`847{7F)5_Vu`D$AKDNVe>4oW>OaePu^aEmd*6
z=DxnA^lNq#7-`J5
ztErBktg;ArqB&>(TugL&TKe#~#yQEGzp|Nm9vSHqd{2qVbrB
zVStlqm&9~=b#B+qHL%>zxDa#2bRlE?@2E1gh<0P(+!C-_C)KcH@@F_>eHh|
z^4gi382cUX+I>zBoNXu9a8mBi`+2@VCRynAFc6t1TE*>7S``ol7zoT~kvhbIPI6AM
zBBO!za6OQ4axOtlJ4cM|y3#WFwiS99{Hg({n22Sl0@F|jBa=&FHE+^RQ+)xW&^9EF
zYW!j(I~Gt;&WuA;>4DnJ12aP}>{8dwH-HQP-`wIq~40Gmx@W
z>J(PVps}SShYpi=P_P^j@R<346A#luvX}@O=)#w^{kKqkH5`}B(b!@jqDlCI`O^TK
zEo47N$`i@TEZfDpJx64d(41rwYn!kB=g+an$|cPgHc)5{`gVv0E->(4^$>tT!nbhxPbzr
z!SlM0`vtB{vR}{JcTAVl-iv^23;J{kX!Gtj07WPsWiGTp$=hWCc35Ng=CePo`O?l8
zTTu*h3mhK3?W*abRyr#*C{ourW(ae%C)NsSr=)7NSnV}Qf#}B4`7eE$vD4fX_}tG{
zHM0YIU6*}W_;MyQX|1$92eQzwMEyyzxrQ2&Jdxlh-W@b2F+6DkrXS-w>yds%Q
zA-*?cV{?E-uxAr)n`g|Cgfxj$1+N2D1h^rt4Ee|MJHBZ<(4;CwD-DO3)x%OJr{Z6n
z_aF`7hSmmpuVrsCueI7CikfL=`G%Y8zfTwkbjvMZb^7&bn`C-M^V=}kqtrf}dRcq2
zSpvTa4%LKdHWhhCx!eJ;>*@Cn>wU`ZX~G4;O-jvEj7_*Gly(!0kJyw^y8eF-zKDQ*
zlth5a2#!hY#2MZ-?<-^B*&(h}O;ihiv(0L#jZx{>E?OIKMdMFTBD-ooFO3uK%A2?z
zARwCzzbgf=IqFs+tAMg1Sj=>U9!W^LhUfuyV?6Q(w4+`kh=Fap*Sa`
zzL?vdu}fdFu5>!=HLh=q5$*w-e7LnfrqSlVi=X=KY~2fcufI~=2Z1&XVN3hI$E9sNlnk$9r(D96Tw0BvyhJV3M~3c>~Y$tCjQheoJZSr7stn
zJwOM7bw1!V*yfe>63P-AeO1uPWs7o^%o1KGE<#r_(B-H^9|?H`wV9~$4thIIg>8Pm
z!0;_mgxG3@W(>C`JdgwtFHG?P3ymMQxGmQo1T-;<2@K7}X#i^_kq<^n&c*l&YX~j)
z9*dP4KrRbJ1|}-2`Uk|61ZobeH9XwLzbQiZZs4XA92nH0>z5)&ZTq>y9M$_V*-?SnFynVt<0F>K=%((gi?=*BJ;edXKA%144y`9pb
z>MW><8!8k-+#)TM0JkD+WFR(E6c{F?fuukgD`9^|4=j}aKO+7tAk0|QY=frvf6U>?
z$(TfD!pa0mxk86=PPX#5H!
zOp({d_eKw2FF%Y`-GA#k^as+?Oe-s4BCn2}>U9ljN*Z<3d^1}Y8>Uu+^S&+-b0-NI
zC#}*WyJw`Hy|j`J4unl>*}d2GubyTK00?GnA-|IL(M}Z2MZBnPr%n|it{|~nvL!q{
zR)k7{L;W;H<{DcskFQu5hrOd-mTls{G5`$2q+q=}wN5!>a{T>Zkt~c8@DkYe{Np>v
zna6IwCOJESS}Nbkdvg>F=ma8P_{b13p~Cs|GS6}Bm;;&;D~YbkBABj`5GJIS0R!TC
zZnGhX+D=#(-5c
zb4?jjE-!t$7mN8#1?0Q|piS>ZPQqyB|6no?kZ(brY2honVtfKxhqw)B(d
z(*K#`ZyHW9F*^wwP8*$i=o%)kTHPI9Z!?C}Ip#C4Jpdqq;ha^;*`z3Nz+6E8C5Bb4{~qQ{&@K#5X^DU5
zjG&lw0heNn`2#qyhqQMUl(T=LD#}mDy1qu
zL8vwaEpR}y|GT&sgK%8X99#0kdfE%>xw2@R?s{e1{<*gy46xp!pg<%${3U?!V=k{T
zg0$59j$gx`-eX{!@&YM>Jzp1jbI|^0RR4@cOJ~o{loh@Y8Wf^gKlXU3?C`Uf>NiY*litCq~fTHE9Oi2rdqPi
zF`{g!cf}#%`rsfJEp=WnB$L1nAy4wiEf1@VMbtG@3E=7L{n>8hI9c005xly!sMOLS
z??jtrZ0*bTR?O2Vt2?aas+-*o{5gz3zGnRLGkj<3>B`Y;#Q%25md1Lw=H8;aNj@mh
zHFLkzG)%ffc*M#^l7YMA=?bM&`UX2FT_R)TEwmAN-Q`@c3+2zLcM(
zpMnt%N}M`IoG`abtv<{RmpqW?kE9F?xzjU`0YZky?qUHHH67dz=Y?qL4U*8~uPf51
z@f5`cok6MeRqLERx_G4@f->@MP5NIg+JVWFj|!emc}6E|)4rXZTLiWAQ2NYj=a6nI
z6GaNE*5t27Q8Ol*mu2kh#@QwzAJHLL*jd>#ZdLwmjOFlR1YR2jw?;U6F=P;vvZp+`
zE_KO17pBxzb!l{=yY`W$6Y#BMI$Rn7h@9V)JE{6%m2c;Uw2uyNUhjE!8saR5zFX+$
zX%DyA=zO>EU5l`8Sgm+`5Ae7fogTj?cG%!xoLY0|WpTv%F3_Yf%2k%{#9L>zDL%5o
z1v?be2JAR~(yry3OGYPbdX`G9Vm)W2<%mm-k?b;So1qL)4x^fn|4O%t64LuT#}b6s
zJL%|(xV^ml+v&t;l6`@k0^s~O%`V22B{|2}CP7ex8tBUVn4k*L%7!_43isKIX4=W^
zsB`BG@i1br=&M_2@J-#CfS8=tT`(RDs}Rz90!TJOH$Wk|Rw)R?uRr-ki2t+
zd`*_**!%%RuR|z1{tkavY!?m;m!@YT-%GJ49jJ7b|KN`!am0JfTU8&vf!?2t9_N)I
z09GD0ZTivbpxkG(MEc;MWTaPq4pAC8&owVeIaV!@E96j=qca%oBmA)GG)Fj
zppEb5-<6w4vR06G7SzW;9BmwezwOo)mnk$t*tN&>AUhn4a!~$sJg|obUvMx0mk$Y#
zO#BwA1J#U4gIoAKYOZ$$z~-9+Cps^lz~P77NdQflh&=9}pw~9M%r9xhj$WQeOIanJ
zh$S`AA!1$0O9s&QU>T6$qe?@8H4+%Z6^dQt$bHIs_{p%}akTX6x-ia}rYxs-$^F{;
zEbhqMMdD8uk14iMUuJ&a#$7QBaz8Rn{mXL;!uk@idrh3F1F4+W{k};(%N$iO0Db;Q
zScO}RG480?lq@0Frjh*wC-$G~wz_TDR0dSAU}};wXos;)zX&qAf+~>J4;^dYWi<=}*kvYK%5bOFt$e
zckGV)3%!NkdiQ$<=)}hyc76=Ez-6^_t{@$Ol{)wOzQ2=uyxT?W0fP1kckv9aj+cAh
zK`d{%%E=+#PrbE0>omd{Q1j99)cTk$(QoX&)yx$)Zt6IlWcEVo8wvO)kfBg`CntmE
zfYajeSi2SO${EplfM4MGJd4KbuKPD)2r|5*LGWIOXLMwLLAo80Kj+`~fw>tB3uD2a
z;#F_`NX8eqcX~vzHuU6v8bp(yAwkv^NbBj};BpgT-NB@}i_g1A%tHI;^_g(^wAjTW*X`w5pu{g|mijBrvK|DF3)A&%;(gvki4oV_Lfbqnu6J2VU4eY;}gB~*!4HvP1~jYw8pR^5|O0gO*`?07pMI-Xj83)e+;d
z(0o*a&kQj7*Ph#Ino>7Wf&SKfK}DEk`bKTUtIBd;5xE%+pSfy@I;i)Ed%>||hojqC
z6opPXGTI@pA3I1}+-ni9L%711>f4y$s3BoS0~8oWKcdSgwY*4;A(FU_*Q%LAb^ZaR
z7wtb5e+qn;ju*4K11RW(G#H~B=2E$(x%H8)P-GSD5Y8g1O6;hrl
z?ym8hi!*}>KsOVZbCNI2B`{DXG*w|Upjr1+Abt^dN23&1JwC4e-RD)peJxdXCUiSq
z%MKe5Dm`Get!bHpFBNHS;BUJLx(8?PWkDa~{;e;np#M8d0aKzK&15)7I!M3Xb@B%n
z^%uCpiM+ypZt4se?f>BZod3oB>4E>p`eP^rWn+3Y)PVTmwqoIYM_7#|6~TgOmfx5%
zS4BC)I-C0vlr1rrX)useXS}@Ihj++(d-Wc3v$wNZTaG~Q@`rA9SnHcLIxM%hY0>q^W0x0PA_EDX=N+i#m
zb$WE7GrJu9hU~``;r|=S8e$3Sn3yi~>G0O7%O?u{3dE;Kh=8}Q;GXvgq{>Wu6&_9WpLSG6G
zJOAX@_@l}m;Ss866hK0A)>oul&0x`BO=B}fgB=;IOF$qY-MOcXgM<64AaUdBw9o!T
zmR!%Hudo%iohw$kjs~LjB-;9}-&9e2S(2{WdF<1nwLr6$+z65`m5ZGo(W{|>j
zL6jhdNN3T5%r+@Vb5w+vPXIY~*p2I;u?
z0Bna4IW++=oq;XxfQ(DU5d#^Nr5QrD1rY!tEPrRqAgj>ESD^42)k4L#2RIg(H!#+5
z*KM+X;>uR3220^}k6GoWjo_oCW7G55i$sAcYU?Ii*DY8|!2TRg9+^O@BfU9%PDa=Y
zS1?*K{bdz%2oMnD|LsI6r#HNXTl*XAo8&KIWu)O9V0TD+4eZmBIe!luDy{8mJ{v`J
zj?{`Z!#nCyoq)ldHC0M3%WZ1WIqEQVjhLP`+HsyFQ#Fkrf8BF2;1MG;l<1h!eMcyV
zjV3k>eWuy#BE!3)u}e+N&=q~WoS;vSPf_J@r_nQ#1jxj~40q&Wa@q(NytT9Ho+i})
z78<@Iv}_oC`Ma2iQ=4>8ajB`&W18o!z)rv)ZEyVpMf0UtN5ki{j*Qe~m}bvmohyeJ
z5?DA^IgZ?xcHJIJCni)#o<&xf5u*Uvf=-$TQ=$H1ab}joQ5xV2P=pD%L_L<0
z`&wHw^bO+T@iSv9#SwPEFwM;q^M+ubRuqVhECd|l0LWvc9MwSwTYk{GC?RvtdHb#K
zzGf)Css_{`=tf#NzpC#M3Mbigf?vubXOy
z=3D*odFXu8;<|Hr6z!!mcF94ZB$C4Cx5MIGSzq*gF~#FYeG;b;gtV!#iZ&
zL`fDj?{U@;Vzvq-9XNQ50P*pr_GbCjp10+NXiDmb%SgK;tmmJ6ip(q9zEqkdyX}tbZLub}KaCDNQ=3Cc
z(T!^i@xy|%c#m^=hL>u!JNtDyex|W~OqDpZfM@0vwBK%FuN5?W0eAW|>atv`Vj*>nHxHD|$K5
zIMDOK`36)eQq6@5)B0A3(^J>#V(PYZ*CuW+dkqv=i3EOOR7kV0+gol;86b{0;5}og
z7Q7{GCa5bK`(^YDij`T7LZUv-=XvreTt=@P9{+u^Of1-%u;V{TwG4|!3RP^m1Rg42
zSG9hbq|Wb)>o(`zm#jo%I@Lc-S<61>On)q$V2mV`(j_N-k6)-KO4*73N3Xcae%_7z
z@n6w)M;!^wQ|xFILltl+lki2{lsJ(&9}FL^?mZb&vRN#vt-&RGcyAS<8}
zzk(iKSP^B{Y(_COh%`N%TNwp00A4-{Oevf?*mM<>w4T0rQ8L_lZ-#p9C
zuk>ml%Z8hWSe7n*(B4mwn1DvzgBKG?CX!pl}4|`6*hN
z&&DXLei|~D{&Es#%$h}$w&ij#Z&XJ81*#^bt^D6P1rPK8GJpRUD<<*=MGO4j-?;w}
z|6mQlVVE)o@DGK0D@Am35z~W2>jTF2>zZP6M3g)p|A>F4Xv#I2+OaEMdgP=Ev%8zz
z>rL;d^C`j=AGhQ6=h@;Wt9q6)#jt&no80-SenAP)5yf4JNQefuY+hn0pKN5IxIjL)
zq!Zqah+IB~iEF2s0)US-?~=lnpGmN@7yh474B}9ZpeR#yus4xjX4Ibb8)c3?@09x)
zI&f3OkwaBhZ3qS&u};>G)g2vHk@otwr>ljUa>M4g`h3IcHa_4$kEZNB?Nd5;g;r|l
z&faFN{iOA75n<@$XCmqsz-ouS)f72e38^}|Zj$Fb-i+#p0VF8C|8Tz^Qp6AIBmTR$
zVS}G4@Oxw|^g3EWZB}f_dhCc)UV2!$(S7FG#LGOT>emVK_s?YI!{0wqQ6qMh=s1L)
zLw4Sj27VoU-~llf#y=6_aO$^z3CSvg3Lcm~k-k3c-Iq(ok-XUsweGM`U8PNX#%QGh
zg95M5QfS=<*g&ovUQu5{{|sofsotB__`7KZwYgC|*|RLh2NwF#lB{`n9SMnkzWdlMSn2x~=-G9cRM_93(Fp%N5nn
z`JtG1RwvNMp)(bi(Gz9A8(NDPML!C1#O4%VV}Z1y)w`OES*vH5zm(?RGPG)lXhD2>-#~2F
zL$`P0DpzR6H4@gizX?o8eFO!;4eifnTV1UUcE_^LGJ?itoORE_?^_j8R1x0JTTdb!
z|2#Ct;b%WGs+Vs5t!mfM6q|%W^U4~(NM-!ZI)-Ukh}rCCDt^AM89Q72Cn{AjFhJ%@
z4SM!2ocDy7jKly-^Su(TtlieKXQe=ekuuG+%_cAH74qg|6LP{3pj(0_q8COe^Q~VB
z%^@NwaSCfjlPqZZq?EZ2f$c9pM*pdK;06rRHQcjt^!uJdgQpAQVW-yuKTvFI5RS7ZJ)gPnY1R
zm_A40@nq`ZB+S&+uF{d$#MO7X{rFR3nQ0zPdkX!1n7;JFrk+nN~M?*-7_c2SC@vLhMJjJICu!n7UPdqM`)nrvp!^#S&C)
zhp@~)tXTItDElSXTNfQRpU|x9I;a<9E@6Tp;R?eQTC^3LMu&e$k5e^+T
z4Hd;eD))Tqdy91W*=z1yF!XZ$Si9H4mEa7LCB$>CPp5-J?UXN;<;N*?Bl4tGDMuKd
zw7zWe=vYeYRh+x?(+grBmcnoUd5Lg%{x=eglbPp#kaj$~e}kbLzrgUsmI}mwV00TC
z#$a@vL`)$#C-_Dp6coenRy-XH8s4bqb1(m~bbIOsqP2BWVpAFFyKS#-zF}6Z?-u^t
zO%IQkFBeOGVFcvtR?^WlR8PFbBma1c5O*MrM*39VJGL4){okwA0g#-0n5;HY{RE
zO;08W34s9fDG(A%Ht?2thhImMz-}>7w|b_l94G5=e2R
zg$T9vryB_No`>BC-wwzg;*MWI7>z(POxJ`|PcJx)B9>!Q4F#b^?KY(Jmdo%+rDGmg
zp|4I_MfMuyB%l4R<*~yeNyHk80xGC&C-vvCYxW1VR@h!qhVa;(q9NRIdH9XWtyklS
zF<>U!Nbce!BCb&9LaSVqfp{aMbJdsiZ@R*u8m`Uw<5oejb7)B>WZ_H;?cR*1t|{~^
zm%IGJ@EYs{w%@1Oi&EEjuGtdAH9Z!t*`ff7XFzz%D5a*s5J}rz!Q2To*}ex1`s9!l
z*=I{PoScifwy*Z+Lha!+l|OLF7Gq|~5J(6<^Vq_uT(SaO$Ul}$Q^1e|D
z@C-OJ!Jy
z4N((yGfwN`lRg3DPXigirik-0Qr!st2}iTp@{>%k%{Q3QPKlDI!C1CKrQPYIl6>l58q#t
zmD~k{az9@nbIkB=<+Luv6md-blZ3~bINOYaO6BnQPp9#?{1#@S+rEVza3>P?-Wdd5
zA&1O_3k>OlTR*om{zFqulkhXjgl!>rLDk8
z$FF$UJeth>)Qt?+w+e84V1`B4Yb)uD?0P=i@q5$o{P!-a6wh2>tckmV^3w7?LGh2*
zFS$oz?_WeY1m8m6vU3P*ww?yd0_FK-V^-wae7Q4GXfL#fiOyE+qCPe>MjMWMkS$^l
zl~HBgHn=<#_?7-axIG(&>Ph9@dV5Xz(Vc4fo_+d*xC>kU8vF0;0H$t4t&5mo;^W-O
zYKVnW^J-{w?RdauE
z!`M@8HrN6K=yif;ocKO4dQK-=LN+-&%w5yP+g=M8sJ7m!hR?0*(_
z*1+EtlSP7A7m)8fS)AMH7TvZW><+jT)Td|kK_1kIk1>q{&WF0oyA1IWp3LoNO49NX
z_oR5c;m5*#&a{H%_eZi0?WmJcz8vH~GeKv;*ndx!izT`ULn0W49qD!?H2ynHS<$0$
zokq$!{gE1Ok{pD+6ch567@82U?{mVOtf6C?$#WdDD<}%_mk6r#Fpl3n*-7MsxiPi0
zic*3o(9x%HB-To9|_B?B2O0iYPztp9)e!^*?T
z{a-Ob0PeW!!N{YaQNLaREn=+WMMud+cI?ErQD`YfK8q6&RE0)4lI#((>6i2SR!6YJ
zQ~3E0U%$5zbFWRxt_b8P|6hLZH`7)cSAW_W-5dP)FVu6bwPihR%NN)gV{Ek!Eoy%I
zKYMR~U@1_$J+mlyEV@f^t)97$*x==5C++zGUcFN(Ha=7vzfZZ@YdJph>WZRBuYRsI
zdL0ZE3~pXipZn(b4A6C(F^^idtEuqa`MY|3`Ms|&MxEQnxTTMFnx-l=jauCz8AxVW
zTF+Rrf-6>!J-u7d8_(sydi~ly<8>Xp7w9b^gRfWYHX)kEnTd6!P1KS~G4g2qa(zYx
z`WQCn(=ZVO8&5&E!(-0Q7TlgvaG&Wkzw!^q2dQ5)
z{B3imAQgw-w?+4LFf(MY;cEAr52=AGWXH^7BIoPg?SOyDU0@){$Ec2=NESZ*2;#=Z
z5;8Soi3t&~Al+~mrTyNHoik$+3Ol_3blQRUBLfdL2eV9jle$crWJW(BG+Sf4I#vtP
z@csII>=1aD_6^93fJ~DK2G8KT`Nr}?ccoooj?wLCU%W{yXcK)xZ(23}HdOueRv2xu
zE#QJD;J>d>0;925gtbXRT|dc9mzg~?SIiX2o|H)8YE@JeZ$laeWeCr`45kVgs|Xlx
z;5*5Kko9mrd6PXR+8NPH5S_@3ERx#F6RJ8>fT^+Iy(Igt%+I4vzO)^c&U>&egp!SUJa{*fJGHuwx?PWj
z_~j>umm3m+oYH6NvQ39hiA)0^M^1uvJrAKF{$t61Fd62=J#otv7K`V{VWW2N>|
z)5eS`&~U1PWM1OW1-hDaloPk=Yol&ij}!N=_a+^bf5(~}dw51T>*oNHOzG!@U99(IW>7XStk2SV2Ce$@U6y*7{ReEiA1<(&AL_i!ChS
z83dclCB~qKZpEJLj3VIJ;wGBw0V_mP{}5k&Dnm|61=diwO9H!9Wtx8SQZB?WO08BzV>3E}n!2ck4Ib_p3*lVw(avz%F-FzC}gGkpjF
z)Rm7lFD|)P06{%-e8N8C=VkXAu>!IqNFEW+oxq?jcd~ZUSt!~G9p6y1!s+#A&5;Oj~S5#
zbXUeHEb}vaGISuRPK*E5uG3#Ltn!$KN5WAU94j3>v33BqHKemBCgjD8n&6~nITC}h
zeVZ9W03A#kf*VJkturu{joKxueW7TF4Ljf4;rJFq6@J+=%ypf4tYur(Ws{%7+JS*&
zew#H9AumZA$qkGE`TI*vko8wO=!KeSe$p@}D-Ty~67<`@R#BV=MyoUuk_ehDmqgVC3uvTSdv_;o
zB?tfnwZk`*&07}9R$X-y5-r>Qx5@Y`^^PZbu&mt8X~(P#;;1^i>T9ssjZMSde$s*dsL3nVs%uh65XL)>m$MYvcdUCS&c&@fGP`8GhXss{Z!`dVaN!6hahY;3_gTaGI$4z_7*K%9pk?7zCtkLzSt4pd-%$8
z1siqfKu0Nwjbi0$ryF>E404OYFRp7(GxnF2GghUA#oOwbUWuw5_1sg4qdwf=dsl$%
zW$SwvP>BZjDo3UuSEni(*p5
z!9Eucm(BGgZN}7U2SKl2I#^GT2pSLaN!$98z+RxRhqYuBA~QeCS+tHe*tnpk{p6wg
z8TU5{YtqXUUScXXX!ecTT;L3o4jG6p)ry8}LX9z$_fTh_$cp;4mpm>bu6>d-CSUQR
z1ZG~<<+gZ}`zBeYZIJ>tbr9{1awH^^QZ|f8yurq!h%s)H-&kA7YrdaEn
zL)xGoYlU{B9%3~h{q1^jkyWhpNoIVSR2!Gi$o#Lm`jeu_>)6${+c;hoPwhflx%7_7
z-4e0hp4Frb#GOkDh%qDfFFpPm_y{Q@W?@rw*)VMRXGq1Gc1^6|vPO{Z3g0y8j!FdP
z@HN`f<)8VoKQ5uqSOj8Pj_m==NygD+fH_%h;8Zmqet@+*nycyt|;!OD|(02
zIm_MsaxST&+<|e#j5>`d;elLMo6;gXJ}IO(dR-s9McHfPV*}7fCiF6#K3L#=*k2~n
zEP4s!si`Cra`WlH;@$!;wgiTG%PXFqKylM?>AVDlJSB{kw5?Y7U*cs@>_W4-=fFQ3
z>KvQ7XU6fW{#*9!rmA;3dTWLwr3Na9u~=w-)e3Cc*-O_e07jBh^HLrny{%ke8!5QJ
z_zd&+``Yg0%$&d^Cn(hWU{ggL4$mLtbL05wiAulnwGpl?rh*Ledcl-aO1Pud$?Oox
zw${PGHTN!YwX}E*>k>dsp(HEQb<#mwyJfj4Rrh?sWqJW}Da-2wrgSd(%tOBT${ea;
zJp3}0;AaRV1rq6ieFGzDXCz@LzEHSn4I6Y6j;fE9T)pS?y1By|@{qOH0f3sO)u4DM
zG0IUs)Q6O&F=zwGh>CdYBa
z&*kAXOE-b^%4$*mcQw^i7o@_rs)NO6=mp=F=N}2gpODCiFn9a^Ez|!|NEqPUJpWUc
zr|IYuD?uGD98xdj_k;?RLrZk0EcGkA({
zx1J0dRgzQtQ{lM<$QfyCD?KH^%ZclDNGzHBgQZ{d?zF>^*kuf?Fn~)S-+TMsD%~F|
zJkMj&Xm3(HfKvUuf|ypYy5{W<47k}=iwny0yRbcx6)#UoRb(xe%~$boZKY*HxriIH
za+~o5bxEi>l$y9`RR(3$LUcK!U?%_UCbqVgD^j?}$RpoR4|e0@7e;C226Ifjr;i&!1v@(6g(X;P3mc(C^v}sR!C@Y~-xJ
zllq!YO>7ft7Wkl3gxJslkHIDq3f&CJ6Gxm54--B}sgmNdeGh7*yAtB>XY$u!o(E0D
z9cn^1(ZuuEb*-?MovUa+f?4hRhu~+yRYD;=-}ARQ
zhZ}718gTOAY$1~fS(Mu;;gMNr7K_3F!oJ6_ON7A1xQ+@?lfP1KPp{>XDGaI
z|2bIn=8?};!~0MNK0d4*VIwM?Ut|m1sA7KWyQ8xMG0nD(w$0fTdE-L9th59ol?QDX
z6coY8eW*WU70X1;Pqx8(=gkAjx0K0)>DdU5BbVft`aX=()DByJ4kl44t1?qo%x;nF
zqG4OfG=z|yjzK>y{Ze?OseA87W*s*f7azSs5UCI8;&{CU+KAD)HAlnUVE2hUKtkj0
z31%4gE|e&qrzC^|NIJymU{riR^|_e*7chg?PY3>lPLuF+PmAC8%R+{W+7&C~=@~%v
zn1nD}(TZ|&uX$OmMH4ePVwZRxf2wVej(9tEoLlVRTrnE0M#19_cC8q;E09uQpkYi@
zmOE&QK%55vTcXJH6Cq_0SNeP6pihYAiQs#Jo>AVWXaQilL`0wsp`O_D9e0So+S4^Y765j_4*R
zdQwWqGRP}s2;Vwk`A7nZi`eNQ>KGucav+=_Yg3#6B5#&GaGWG`-z$bsP_v!iUSc>{
zS^J7R@C+75lN&4iK^%AlcugvZXuslHoM?Bbs=`EGiqRsQ9elZIc-k>lFy~b3u_iTkB75Y_a_er$2U~(yN!l=~cG=s4lrKTbcI5#SMkg(#9$;U6H(M*Hk-b=sVORifY
zxm#X>TQMOcYdz4gGLzP_*`%&vUq)&RXf5Fc#=P
z819OwTqf6&-O#QCbn&pn_q8-K@t>uL7Uc#483I3R}8VGc>Quo
z!$gcloGRamGF{F5q{xbY(MWVS0do<>--8*OpL0ROlLG=H^kGAPwIuDIU?`$me7w@`Al2t2sd
zWtJmRMnMaEiwX+kmsXqKNU-#0es~r9QOpg0An`1m8rY={_lDn&LBpYw!9_sAJbC5J
zn^$lKXCA6pD>FJaV{RqG^F7o6E^}FB$%aQ;bIjKCAi1pDhRPvHxS}dZCXt?cOO++EsBmkr-os*kYd9d%QNZ+YmM@^E2B>EBr@bh7<^@LORptg
z(1H`5h+-4|y5~ig*0nsq5Aa)8aW^$BzW;d|U8oG*lxlGw2@*xLNvkFY+~a>=G4@5;
z)|SR03GPAKt(P=p|I031K+Jr|gZ`Kj20K7%%wsqHnobMJFH#jmsBhHw$!wC@k!NSX
zz@#nA>yDL(ly}8?XDDmx66O<}go2x=6LO6-rFc`_biNw%exgFX1YY?xW8sv!{|p*;
zPDlZ+Ek$WchQmu4-gYesTnQYc8j+|AX6#E%ywY&LQRkCc8;$es%w(a^5;v6;oW=ag
zG1BO?V4lCh^j~~)gM=bkH616|^-D*i^Z1-jEAVKLB?0?Xx!EnZ+`RvEZBe0OMDq%G
zM7wEKw5M9LMlVmaF}ggE-CMS*_o8H!G}CudH7cH|_}BR8Z3mnKXemSC44iGJPAulX
zaekgWGIDAppT@;Mx#ZGFN1quLPKNxR@962>(=%VgK)WKW8bE&~tD+J1l%0SfUggmp
zUPjs(UN2-@*?ieSd5<|IgRO^C1pT_(Thx^?iEyK+7DbeGwf6UEyQh5hA5Rj{B3+Im
z58y)ZQmv|*ENyKC9v3>tKG&*)0(8d;3Hlex;E1YM?S1TA*3OfC82HRB;Xlg+V@743&J?DNKSW1Ui!g~=qT$h
zcr%uQQuQrYI2fH-wZYCJcFRO1B>lBH+GLn`Z4ZIL#H>1?M^Euc`rY*%vOcon%gsO&
ze-^IlzA-tk^9HL5Ptm)K!+xwmK
z1@?fj>Gq$x8y$>=GiAF>5XjRRPd;o#?U~hDOd>2CK_O{>rHbHK<}SJIC)^7ig?7iR
zik_!$94htGHLBkfZp2a?6GWA^&k{bKgl66*;JiJ)CN_Qjd^{T*RpQ-At(uZ=^;vf^
z)xes=aC(rWj+$sPO)2eQmFe{1b@)U_%&X3>;^F6~(v{-l;-(*-2R1^#kua
zG_O15z-G%>*#w5^+uu$
zF`jrsV%2}!tH$P-0M)a>il@GNkN>ONmF{Mm)3yFqgk!}_w$6=$JU7kGzr^_QGL8am
zj(3NyJZ#iFRDY!_$_qZ(SN9w3O?cY$L$z7Y)WZAtJ;oBMzw~sP0zlhVd#yf$8c`nm
zF9xjj@e&2PxKn=9o!Vf@Y6`12{m`b}D+-y#tJg`ZaL5Kh0G2VDRd?c!$MDhXvG9_X
zw-%ky21~#soS{Jl(&F2Mz23^DF`n~$_L0DoO-a|(*Ev%(ypq(;JLF}y==DLlIMco&
zw3362$mcyJO5?Dz0!wr-X_WS7OR>i{Ve`|Ny4X>c)+j|
zGAe3-WBZD8O-aDHvy)a(zvZsd(ZzRy)o%#N?f#bKh7kY8PgO)DKWwn%O0>ofln>h#
z0dSNO_zb;1xy~T$O>p_|GH|cef9fB0BTG;nf^f+ANd3xJ(o3^wMgKa!uR>bR(J$I}
zVVmGS@3J$)?!X6tG}&cYz(LJ{ni|lIeFP-xlrXbUB5`6EpiBEX#I#h=AZ&y%zk?5^
zPARj>(Le@6WOtTcpry%*jJjIJYGX}L0JvCQy+T@{&BTPP&01@jHuD@b^gb$w+|5_n
z&{UFXk~2|@(cspjUc!8gMSY>cCnU&@9ahl&>X$o%FVRoZ4P9I^3RQ#g_(z%YFSpE*-h34uT`O47-++*_9b~xb5BnbM44Hr)F=?l
zq!$|LJePB>A!%R`LfMJ?v5NBnEn+0zeAFBm0&zfYom>*GiFrqG!=`vAA2SN@Wm@4Z#MiW=Sj
z@^?vsdyRfqa(9#$=wHE)p-XjYJ}wH)e2PE)@pq}>j!tpvFCmz&9wgg8-LnmW
z?EY43Lrg78-S`?=rm42euCc2*gjqv{wY}_a(JwrEfBNICpDXf^695T|J4y}(rBAa~4Fxvl-rbTS+p--s)H2^k+sEB7#$MM2es@a*?fhenthnR<5!uIx+3#>btL$uXL9>FyJSipyLL-f?iwr8+^66O@5
zbAhxe9*=y18k@Y(wz!yX5YpNRPO-giBO}=H%TH#eh?bczPR%@imBhyi6GC1kt)$p}
z!opx8`slP?{i;Kc!_$|Z-!OCYG2`jj-Nhd1<@m*BX(Q+l=k-)RYy
zP3NmaO?k+9BLSsTFLEBC#TiG`&+ZJC?M>tL+ru#RH>bT!s3^CAQ*RsAmFg6}0#-Nw
z&owPcgS#Wk%QTlQjW3DK0eWqCipp=ha9zeIl4#7OEQ(NAHTb;1U2b3*S6oQGV$>a{
zfUyu`!BsoIZC(^ZUzkR+MMi64bsa%1lyyuTcS-kw1@If)j7SBFAP0OQ%By=8rMQG^
z)9Z>ZuQyWtXRadKfSZC%nOp4SjiJ
zO3|ynO`7Iie#bU#I*$K~$=IcH7t)Av*k--~3ag!$o`R(>%>w4TpK&H4tB1inhxry-
zkU=fw5`a=BwvSm8jOGAMu~j-6l-UTX^F6>k^|Ti!ypXA~`^X8#U-5YWDU93twHBR`
zJe|008KbRFx6v8m4$Xw#1O%FG+uf#KMD`s26EF_?dgDqW-e@@S0sNpBS1HP=k&*;6wRZ?{LR_e&8P`7|NR?*aSUF+Hs=;IK
zYHVMoP%A_g+%Q%r9nZH8NgB>?1!Z!v(zDe`efwB>brQvyU=EVNY@NW5q{cZMaU48z8W#UNtV-
z^L{^|E@Z0TFYQyaAdRjzKWQ(jQ5ml|_~PescuBUdzj=onpbn7|)Jh
z&j<}_sQBj`{jYLrx1C3?#v+50y8(OEv4&hmUCeL_Lrg@KMg439IYK&dd==J(+nppE
zw)6@9AMG#M>iHMa9|krA_JDJF_zccY@zZJ1<-xH7C#n>>4QJOc8`JbqM~Ob
zj!9Rh1znG@jo@=&q`O48Doj=Qr7)CX5P(b7mTL{!*Scn2Mh;Vdiut18(@`ZORmCI)
zN-q29CG!`PNKbQ|XRE>LmuUWpFx>$Q$(E=viyYDI0JG+MIMKq6o^E$;(>jHdr*HKJw!N+4Ib%j8#1w?S)L{!z
zKNvE1?l3bPr_KTD`6~~EiYmzVELtp%yz)>KTsqYTtF};&Tl}yp1e-xt4UU$u{+f@h6
zI;Z+MGdQ?ZMG}Em8fWwMjIo-X)(ggGUcO;GkYI1AiWvfZr>;SNc1Eg_mr`Ck^WqQC
zE>}p&yNr1|+;m(*KfXG=7e70t95(EJ%mKaaT2F9K@R?w
zI%QS2_)~$g%aBbXZ)fAb23jP*pu)59(E;d9Py8Ihm;16fDwOtF2GL(v5UgT
zT(*+}ezI^9u27yj!Y$kP5L(IGHut~_9yHLyY#i5ASO%!?f5*Dk@Q|so;^rD1{1*UT
zsUDz+k_nS7P`+oM>bX_aVL8V06Q})D<=XKm4qt@=x*cWC<$~2@dJ3=tpO_L
z?(czrt^jscRz3tOfDFhH1|9QzVhs;xXD8SH<07GzdO1E-96j@rhnPqUmy#be-Q)$
zBs^SQ|I|?W_mJy<%KV$UxYPU04DJ1Y0)SrsyJA2`4>#X`X!GCGHg|G#vvG5G`&UK~
zz|zJZ^oPCMpEa{_{L7{!rYt8drJ>2B_}+MqOiE7g>NqmHd%OST{U@K8q#{3nmz5j9
z#>WX@eQ#1JM+*rj2ZwjoZU}$mleBr)$=%7-hvk2|+Ro9*%hB(Dt(G>97M6d1RJZVO
zX3=uAaq$4jN&cVjy9wcMn-$0%zzP7l06^a6)+~QY{naUd%xr(m?-B&~IXgK6EP?iJ
zpa2_7(EA&LpBvB<1aNot00sE{bKt)b0viv&!p7YFy~*D<3c_F6PpC9{ZnK(jNBq%0LGY?SBjUzg|EG
z8+)Js_WK_f9nc@DwEvG_8#ieiZ;*wmjk~$^zf}E~UCtf&-t=OQR`#Ix75U4g^=BX1
zzc=jrp=0ys%mOg6adQ99*n2OU+c|>V+yLDF3In}s@IQRth5y4Az#^`HBrhqh$nZby
z@|Tm8qq&oXjiVKSor@a)bae&#Ah5po06P~Kz>n>{l`TNte>Dk!h1t=`{XGQW?BN~&
zuyk@o__L7QTmTlaKc>GB4?BQG_HV=qV3GeDaRXQs|3*9j7Nx%tFMvh)KZy4|i~8S)
z1Hhv3H+s*h`5(mho(1@SH+s)v_BY}Ku$cb`alMb3JK4W$`ELs+=O2rM!{702tgHYQ
zi+?~i01N0pkmH@lyAbw3hktmyGqwC1vcH$KwE4&Idx+IP;Cn^uzw5s%XYJ!`4RZX4
z$Gh#H@$Zt^{sX?x*6ttheM+ko}`gkgKPzyJOFlhJ<_
zp!)l8{71C7xRbYk9}_1pH-L$qj|;%Y{k{h8(r^U)SFHKJt~7t0ZtvUh-}uil4*-F@
zLFNcc^G@ah;kH>Vk(Gf`A1AAzsrZ;rYjA{hM`L)wCKsY{&{
zW%-Q)GeR88b*aL!?O)g1(`zSgE!0GJfPqSZgeX#CqlMak%vzzjN{hi&-Q;vf@`WXO
zpE##;zovX80cedLOYn7mdq`ma`G$U-Np4v6b+s7Ut8A9770cBg!+Q~ppo?&EvJ2e(
z?Exz(9@rq9RxEp5DlQrSYS6AdJLxiMEG-sny9Wao7M%oQV$M
znnt3ZA1#gG32(O%cAdKujn;}U5`%Ol{Qz#9G5p0|$-9i`k?Mh_^sJIXX7>dD
zByFC5rICe8olXSbiIlJ2@K8$rv?cvLjrlLUBu#O&=pP$mxD`Gu#zCLjGWgt!bx7ru
zcvj31zOfUn-JvsF_Vw<8zQ>dnta!VX6cr!>xDNMrjwm1OvYOIDFUT0C^-+ixuGpLA
zKln^W%r-Xtt`5mr&>ZY@E5?6%seHiPMqitM_?dxROEeV9%=U8muvgb}tY`+LbLcn_}K*Jdd%5a?gq1_<_
zFtLl&TKM4mRoTPR*qr~<5*{S;a(2C2t2IlsbCr>n0y9t}?G{G7G(QcI`?24&cP
z?b^y5TE6YT(nqbmyRJ$h*wZ;&rch|*5n)d4oHpzmS^LCcD$-ctIqdu$QCeT)`g$$|
z<7MkZdns%IQgVc$8f76PxoFENp)o5`5N|ll?vPY^^6uS_ugz?yZ0f}323XjS%=)7;ajbZhG8U5~FqyIKqacNLT-o`Fqf-2liRC=e~lKBY+_fx(RsKTl0ziHI^ib(!_Q8CO-DZ
z!l|x>?HVwV8`YMwh1Qd_G>R5cOKD@UYL(1yRgAImF|*2b#ia%$S|$DRGzYFX
zC6OtWY|GUgyE2~z@xnb0Z@vRZlQ-GVI_M%HE8%spOsD36a&
zB!2dV(32
z&c3|l;?*aWD;rsYK)j9v|Xulylj0eNMahz%!SGa1Cf%W%_eq9%3Gym(Jz
z6~&%wO{GAu?_eBQ;;^ggPLy9Fzj-7*RSeV$xl#M?q<*C3=@wlqSFr}
z*TdJ^V*aLzgQa6xu;@y2G8jpM^753l&3Njw%fZajFp@Gg4g}Z}!~{H96Ne_v2HG`H
z5Zy`Wzpv5T976+tiPiYdVlf!oKz=k4S5T3|<{I!#){7pBIxId$dn-X2N>&pOax%yH
z>_XrUF#>AYVKvLpymrT9?6e#PE#BxnFj8|LV*n_Q4#fd9=385=ICA34CeBJbM%hB7
z>#+hxMJPtd~1(0`2FC?|`O_yoT{&A9>L4;;QgfDISsndqn3tNxBKvXF)?oHNk$Anlvo
zK~D$~J54!z_~H~dL!}0}NYQ5J!Z?>S{3dw7RNn`C&o~~cdixb>4msDsw;^=?zOH(o
zDUl>v$vq8!cksuqTL{6J?3XQjRr^cG(8M5(24#5bPg*51`qq-bm~Q($-Ol3%DsI;$
zlNJ4gWVLyHVUJ4~jv5>eSvrgKMGVJt4Qj(ANR;^-?*XCigK{aPuut4F`PYtd-!Vwg
zT^REl)}RR@U>p@=!*C;H5%{fE$sHToEogjHzDy&3=c^d+|GXU?_tfy_SHlmcePfi)
zMimbwKpdIo(_3tJjqdfr44GN}^+onF`{#DCj-7@uJc%rzmt{AQP1Cyo;+0#FEk&5p
zcu&)4p=l^q#7DYM*XpKhELxLZ$)BEYS1s`XLSRJlyB6Q{6Nta;>xsTS_{Pfc>9UWA
z#_f-P;EFl@{NE(tTv4alOUC6Xv~0^GWb9o)Jr#?@%{zUhEJ%r5SK;=(jsHZ9YKC}FK)0>?5mSD2;00C#tm
z4x7&*NH1_&I2K}>YjZG)mbI)MV!|F)Wk}{_W6L8Z#U;K;gQ+d+&6%FEkF}T_$%W!%
zBEWx)vU3n#0kzk}u#!b{3TQw3B7>#bko5F(n^8Md=rvG6_FjFC0Y5n~^+kvrr092l
z+z)0vA~c|R&DxsFIEXqj*E`Rr7~XO&K@gV`B$M8@=xmvi>#M!F1Mj|aoD8004Ox(Q;M!Jy-x
zLsCjK=j2XgAE<4x4-F?N&%2Gg-CYf4Wf9`_1yoY~tFJ?c5t>55bt;zx>_F&$)hbt|
zTo(zDrCU4rI5HNCTwZU@h(4ZWU3kIVm(^`+H~q#+r&3vZ3o|+`flif6{~?&m0_b4rGb
z(SNHX_Nio!+5jcB4~l<@{5tbv1_qe9X*1LEln8y2eEV??)JQUoPpiFulD`_G!F~_{`rzvl
z74L~CMtsdTV+Ohq*+;N`8%Z=X!NWXoblK(eYN=BE>7K9|Wh(Jc@`#2(2ZcW4y$(V;
z@(lIOu+yQ*QrbB56VPdY`m6b{y!EppkNr7GxUGv0N={5f8kDk%Zas#;{sAiG4JBTE
zdAqxhkk@~+i!@?ngVQy|H<1{tT9ERz}
zRGrn8(1d6C*n1c&G4zY_>5wv~8F9xY-YQ}OZ@(dq8aSWhN6ou`6_S~gAo{Gc_HQW`
z$V1VQ1Pfs?^BmNa+cx;{MC&KpCgN>{B`HjccS_h`-47acHqof<6O`&FW~Y)Izl2t*
zImnD51H#iTF!ML%~@of7HML>n}i1f=&HMNE(_^cfiv@b-=v#=|-hawmWmZfcwtxadpii;!=15h4gLZy0*^NH$z0k1Y&E8lE;O?#Nml6$8_~CS2BaQp2cep>h+qD`$c>~i!dKw$g>k9!Siff-_(R!
z#BS-m1GG*IHA7CJ)xhFjON&vVuFG8<+iVngaf3Evk+czyvd%{?!oMTiqg31M3%5gl
znyK4n2P7Z5janVfz?#S#qeJCgETzn
zL8vS|o8~|?_DTrb-R)ylo~j$%Jr!QY|FQOQIPqG(6ZDHwV~7csGRYH@D8eMwRjq6CIRYIR_E7s*KRh#4r7&!1SuxT@Q8&wm8A1fjBzhA5rhqJmb+DZtth^=la4o_
z^Vgs5eJ~!y@jv*syQXur3--JLIAs&6L!jz^*HeJ4B;~4P)6~g9%J213ny2Rt6|mzu
zQO41z_?ZlY6|{nrwqH;Zc5M~BZc=b&lU(~X&RbuaFPSqxY}l`J&-*V;zA5P7qbnkp
zPkpYHJbY;ihEo-DcUj~^VlRfr*d_GpWA2&ye1n<&ML>`!)DE$$I}tO33~FBWaQU)-
z>%#+sP8Q$`{$su+Ps>*^k#O5+zp2Kyf(tio6Cq?KQqK`i=olymx){dj(j+)(G+PatTu#%VYTh!cs!pZTSWW2CK{&iC98N
za3slyg5VM|lKSa=4qd)XiYTw!DKNQzY?&=I#vYGH476zCjQP{$8a}ad6Sul0Rkbb-
z!R&adh>j`n!mtmlYaXaYZsm!eTDnz!cggd8@R31}TERV-vp-(+#;2VqSry#qWDrP^
z*fDP}cz>W^c#eY^ua`K^3X;^I&reE$EycTwT%)nX&@yUMK7f#U6tFzqnw)%p9aF94
zU|fVFapC?R;{GB>c%*kFmZTZxduy-Z_W(B+h6YXc%GannVp747N3*w=PadCkjv
zUloX5BLa7-8>jeW(TV)?6EYuU=G9EMQ+f@w*(J(OdX^Ici>&hs66)xQ15OT9j>VEs
zzX*5PPB&fP%MHZh;oQ$zSwbY7BGaA>h^^4$&j*u}1f3*or~&j0LkBFaN~5Nit~Y9-
z93{HVqU60Z-A#I_rJs<0=zVMkdbgEns?1#NDjBGEX}Al6`BiSv7ZRC8l`OFPtJ)VT
zQz*Wu6R5X+{hClp%zzqdc%^Z_t+Pv%ZNhiM_1tI?x)E{IG5qDM2(6<-gsnYCvJl;W
zcO_*jtBzG!KGeBAn^6ZNZQ;f$fVlr{>0HWrlO=efUP>`Jetk%P0314_t4oPSWtL$b
z>Cz-c5LWs4WG5L*`3H0L)EPs|pz=<9aKyE&RI8tDQy@xP*1>rf6Ow~B-CUC>J}UZ%
zM*sKc=M6(ZWnshwp+L0hCE8aQ)P1uwn;_V7O1w3;o{)JCbdfc%w=5qq$xBWS>q=b2
z)5D-S?6;fS>;&b1J#VhBqOhq8!7_)cf%o$!26HHc?@npKI0OvX9FfLda)t#4vzY+f;p{QS$J;zkWdvwOst_Edjb
zaoQhuXhyPC_c@Da5bT!0_UNNTq1&zxWr-YK#p&i6EKtyX0)L
z>n{VCa!A`X&GGL+`BYJsGN$@ycRL>gOj_2!5G?i2&&6`?jN(sz(WV{g6|>mEw2||r
zG=8S8HfO$U!XyYJ^Sq6Ex}&=CTc)hhKx3;{u^y)-C4qR$nL=6Y
zI0!T=#`C>&-xl4gg!;oa;%ySUGJSIZ*mQMz4fb?_d8ZQfPgyWbPBA#GUwT4VlqG)d
zq0<+pIBt7~dR0bR2{unQ1;)Pm!j_A1V^~9Oesyiz&U+nb^ONtXwm)tQ;c;`gHm3YsxYLsFH;XTV5
zehgX9PI`P8xhMA$6*T(!OmQpZ`6#nBUFsFId0l!pu9n#1PxJ6`L|?s1crvh>fV*LT
z_y-(JNb6-jAak{Ku^
zTwe2`T#(xJ!pGlEJKtqSZ&jX&*UWf-FOlyn%}C2eapWc!eYVkw-4zNdw!q5Gp0FC!
zid1NX&Y=L*aZ$gXzbWDBy_DH+7u|2&wfrraF>793!+lDwYm2CXMz*i`2?h
z0uD%`;6?1xs*q)DC{W<`{ChCjL&6+d@L)8a4QAllK<1_fC)8d~=_ku0W2i`f2UvE~
z)?i4P?}e*=t?{+BSCvHaQE(Et!tm`Zdtk9H^L~_Y{BJINI=3|+*}J~~usj^SI)0Je
z$_zz-$*8p~`tY<6PB@%L`j$oV2CK8R)xPvq0ad)5-$pShVL;`pHmrm`Ff;rY#00RU
z4>s&>S#XCkaLD>@8K?lBwd?AFYLls}NeT`Du
z{C=n~et?kt$HaMBZRz^YQ|7IgaUYjPByH7AYjQfkiN3CM52DqWv5~2N{>+1{G^I
z6DQ?;mgrc062>Jo09|T-nT&}AYM$~_<=P>0(W5cCp;g2bqT5M6Ul
zcQKi^Hr}yF
zl3^}UC8#x0tch#T1;9vW!$5K5j&#z|Egy_UrE%`8kDpSH5W5tAiij{`*f-6PkbHmG
z`XjL5@m2Dxk658`>{l7hIzGB=fpXI&^p|G)wTLFCoXXf&&s-I!_GXn^MDw?!Z{m&`
zv8>d1EJgTPzj4uqW~w4^)%^3agYZ0b;gs@b;Vyz!w?(A&pQ{q9ufFj^IrJ5t0FNIB
zh}YafU<0P6c@&g?fKZ%LLyEge|5<5Uw$Bq+?Duv2gP#S^^#&iCGZC1DQw0D09adkE
z;}7ZH?{Qa^TB!NO_LRHSwXtv4*C`SB$WlqM{Si?R?sQNqD0{bUqBviEsJU@j@A8
z(jd&}+tmtpH1>@6_U}F4^OhQsF|5!+f2*8S&KZcOHZgvpNtkG^`CMtWZJ5Ta?fIK6
zOa;9<7STuziDWX9GxJL2V*@lxfJz`)&7uRX7*|7T9;nP~6aOV0iPWlyp5$Pg0-xFLoA4a&;1jHdG3+HcW
zh_ue7ZpTWpZS^Cl3ysRvjjfKIWoGdXb1A2Wd+_=M3x=f}h^5sIV-vi@>mO{C9NlmMKL~@rSk0l?DEUykV?9A
zznVzHh2D-onFkD#xlRf5RB2e`QspvyXIB(|8t8C<=%N%WmX=0}*}oN|7b@FRJXoM~
z4}M$0ryJFrNY>jq7`QPP8Duo=2Ha2ZX=bqhkl9-bop~wykw|`Il6pdwhre~!PxZt8
zuYoUM-C>;7Tb3U*RwN_yn$fc-7QF~^YNrzO*kg)bT9m2XEx|QVjNWwh%80+28R<=b
zh}_5d{5;Ls?hDw<|0=ydE-k9kD>2G~FQK&Pk~7^eBelUC(sA+Gynp#oT+E#Er@#v`
z7Fl5X^(fipn`zWAYp}bUkx0puJ5UEhwe!@z;q$az5C81J?V;^Y
z3WtSILR3{xYm%fce080ACdQ#vhEOTAl!G+V!_ikza>%`nl|DT6yDEY;dT!&D_(^Qr
zZL|z$ZmGlF3_u%_G8Q&e_9NU}>a#ipU>a3aPix!j9uF{Twn6t<2
z`7%bf9zy;F2KV#UnUKf+lg{Z?nKJNII5%i=a|49IERxEXq-jxU5rP2Z#xe5@GG90;
z$z1+ce~_E~?Rv6Ga~(IW)3BJXIB5rgQQfHx(lt->NR@qvyDX54LM?>!rO0P%$Zs8)
zR}E^o#|`l!eM2a>0SAx0%xv0!uMrYMD{;7%JN3Yq6JKyT{WYDaz*k$j>~l^}`mz;c
zJ`_Oh$ry%LtH#sib!rcu`}0pnZUr+j5{KUb@``{I&&t6QTkP?}xg2t>?=*5esEo^G
zM;c$PgQSGy4R7~}%lfu?9|yc~5n8+pMWWjecBolCb8r13r2h7!KShs!z3O#0Uaj>a
zE92-Cbblo23NfMpmkV@*V0Yp+tP<&b+5qa|5z6`$I<_=U(6&GE!Mf!K4t?w*#5q%d
z0gR{L6^r2XzvyUx$xbgk7|bR0i5rFXw4%2W*jC?H(vhAkB9V!sJ!_EWxb2f{|D@;n
zk%?`)2a%j0RRCuFu0fW6p|>?msWdvty%rfG)j{@AHw31GTy{t)l(W03V@*h{_?V+Y
zOUjW*kbtRa7~3TJ6x}aKhB~0al+#4n*ptMkpQTj>c9LA&rvNE0pg1tC2r;61r;aR6jQYJ2
zS7Oc%at!c>20GdJD?94w;fWPt@<;5REcmJ$tVvn%hYzp
zXqu3%(()8Aezzr9PrE|IFKwE;$}%irx;Yx3#pg+>!In&qhzs`9#B#~$8mChl6$F}}
zNL^0(gu0x6XJ|`ojp)OI%BHh`2rohIoPuB(=P4-PlkjGLs@{2v5+rBCX;gO`|Ea2G
zs{?UUT*XJDR+bh1G&QWvP5Aq_yE<>D1E-}krfG{jign{|&Bam@`5|g78I=2yc4ZL_
zj8+Jm8bq893Exzu{dm%*o->;j82t|-@fYmpfu74>UwwNY7?JhVe%+etTQ1S3A9|3a
zLw(IXWT(7;gBMq|t+BFFT|J-7*>@0+VXcsrpTU+>J!X?cDq;E-s+u`L(8aHOI@3y^
zRBh8AFYBRTEF>>n8ndvT{?H*^C0s=dsdbbxNLA+;KfX2F@;jLXS_Cm`8c}dCs7#`;R$7Z5NuYff)Q~GID={v8hD_`}X04P(l%MC}aQi4_}yAv073;
z&k3-clp|)MoA`A5
zJ=$h_PN;1u2t=#*)@^CwAeK|kJRQsWJ&kE#k|7f~#!X}o{wXUuGSo|5pn}&V
zIu*?b)4d^Ew?Qkp6oplMd;uQ9f&He6>kGxlK5jUKF*k#AL3@43g37NE{?t103EZmX
zXY696)D29Hy_qF!6dinNcumt4-1iBT^RK))p-*11Fa3guI&ByPi}u<8Gl6UmWOcA85W&;)j
zNaF=K+{2Nax_i2`(x%jEXZa<>!8gJ7nE}LsyN~0S@GG-NTPg4b_dh>FIS?
zF({CxHqVZiPi@=xuXgr!8WraZVuNLWFQ>sE^C`pNzHIu(+{Fbk@AAPs%oZu+bn4ex
zFJD==O@&-C23==0Agn&}^a`K54n(IDUC{|gp1b7bR-2j2AUVVLr-=}*Y$IG~=rt_2
z`S=!H3Nsg0e;AEw2Ghtjn1HQ`lj^JBw!lJh{$aM-h%ID=wocg%2AQ0gvT+T6C9FZo
zK%D46p(XdLUV}RY)!+B`ttAOW0414u`b}Op@|7Slq`of!=}d8?BS(NN7MZwJQ7$As
z?cnAP)`Lv-KKdlOt*GcYPzCSA#I|gDFOdsdS4i1!EGH+D8=fwsZM=)ME##a%=>#UV
z%8j3wd}fsTEE#Q8I}T!(l8*_0h$w<)RvkHe0X$Z$rpN~zALrwzbdd2cnbO3TW>;D*
zCii7H!ywrS7VcKds*^p5w9|l|pB}ZOp&2?y9w@5acKd)x4+?ugr&yG>n?UWh
zjR(0xzTRu|Oew7=3B#IMu5x*vtYZ*gg}~?b{3podda&NY>Tne&@T%>vD}qe=WdI%E^3oPKmyl`p#Mj$NDJ$Cep&s>D7~vMt6UANZ0g6h$5EsYpZU(Vxi8
z)|KCL{XcGFV|5vNan*@ByY+4N!#)w3610>w&#aXC6>$BouE)V)#Ssd%RNh?9e$PGXq%bt~a1c+|Y*t0My$
zk-%2U9Q8J7;WxK`DJ2)qvyr|m1NqhYCHQRW8}r8qbIApj()ppZ6V6|58x!s^Uq^c(
z`3Z#M(+q)f{Kl<4_S+a>+mVw#D|WJ3@{zxLV_Tzb(03d$Hkl^^^H+x%bG}NCOZjHF
zG$Ze)(SfUw#kSvnnl&~Vp}!&IQcBabrQ_2#*aw22B816**4OL_6Yvpw*
z;?%@NLFpupLO-NSoSv%fKtQO-;GAK@=cJZ?nol?a}MFItfBMuIDT_=9Cr8?PlyR0UC;B5&Bm-VH-B&sT1{Y;XYdFz_r!_zNSwJA=Hs
z?gxW^Atr`LQkiOZbo>rbg_K-e^3X%IVMdZo&!IPN4ofDTE@Gq6e2614Ya3ZdTmm9y
z2eyyIAnd8a{oQgMj~DRG4luz*aLww6IRU4>Tgn242?>VLBqA-GK_|ccSXvfmk0`N+%0zxzc{OinFLGAR
z@X!PQNE3}0OqY<~Q~;d~I&S&+fy3WPLD+Roq1*mf#=6n~{%9RCYyPs<)#5fXv>Xn%{bsP9^txt)7$qbjbj0&v$s-gH5pLh@VIZe;l99wb>ny~)N
z{?Kc;RCSoVUSpQ|bnY2BO#&9NKXrhJPUKUU(Zs3sY7LL#V3LNK<2dF#+bq%#1G4)s
zYvFWk3vs763eUI15W^P~Zxsn)1+~+Es2^f@B;ibO9VND+k`j}ctZgq{=VZ!ieTk$3
zl(w+YhQJ5lf_vbXxcr4QCVrusF8NUZCLnND06!|u3nU{EaD-;Y&T1CtOZPnavGz?M
zOlp5a&7W-S>&%`t_zKxZa+AK>9gETDrGiiT=!N8=pI^vS7Y~#C1c&7ywH1wj98DgO
z+(N&9D>3l3NFaj=N;0}{9Jc#*8kR2whMD+v8y4L-Mg~JL+@hVk3zIJ70G+_o`o!4H
z4M8Pg{rwRq<9^DVDb42wf{mLP@>|MpIjCVYJEs`ARGm?r<_#}m^)Inq01#VCQ)aBR
z(P+{GdgTuVdiZd$kuP#;QQn7tgN^;uOV>BEEPZr8=~LcRcDefZY!jP(g#e#$u0q7n^>mG~e%!L3gD~Rv4fE%*Q
z1A{C2^yQyPtnQx0j+BWjTfSsv$bO(P^JEAzWy7SiNW~L~DZ8U3)m`&{>s)Be(tf`c
zS|INmPoIM>mqU{QPtL)lK>;&;rPx}|rcXNrfn^O{>sd#LBQo%=)-+Rg$8_j)<$8y5
zRl&)$9+hs*^}jA%`^1TW^PhaStR7Qwx}JMv)0Zd8HMN?Kl3zRXWy+7neVM7lSNhlv
zf8_uwlca1TOxkc%vSPx2gb@v}x#QRyZp}LieB_|0m(h_&A7CA#U$Y;23V>(!dI(R7
zZk9%Xim?q9Ou}+CFxGsiY^EoFQPf&Utik7{1Sam$
z?+V3f569-h-#sL*=jJwqap*@Q)DkNncEtxJbrr6YeShQ79yv6GSLLxs)1y2KZWpHp
zt9xjP^U6vwet*3K&_`DSdsh5S=xk16>!rd?<7!p1*zd!+TFnaMCQoIylJh;JfAP@3
z^XUti=;iYIUQJhjhXiY&zdjMiK9g?(MIWholLa+gZ3F?^YEP53LX*pQFal2)j8q;3
zix|!IHo&vNpZ_^ti{ably;RApoDP<^aNZZ8@;&~itIwf)wz!v35K!A|r0pjol5vUr
z7-x4A3ba1SX=_QKuUZp`Id(X5r$I~uSHuf07hlfQ3dq%eE6|~3Uc5B4os*3aF*STi
ze~UOB@UnB#;kHH{H}WBj^`RC#w2lT5<0>$xXNYJL353sbCvdL~+n>N|tnw^MO*xNc
zPnO2_!z&P3_m;(G8u)#NTzK`Blyxdvf4qKB&kdjozp(?O$qKOJX$TFpctff*!`{h%
zveaDh$ymF8vO}0ZR#=JJLR)Z*brX-bbhybJQi{=Ny7~MVJ0l34JxqWS;oMD$k{f%-
z)4Fz;!wXeL^N6=M>?!*j{w%KIX9Tlgo%GCYD3@`10A33y=Jw0_H+>3NvOQc?j4-RB
zD%W(w^z`b=H119c@V>^RCDyoA0z%Hv%KQly!6a
zBBwQU9VA%WU4AUG5r3j5%TcNQU!4)~t#IJ3Pnk~*keAJ|@~gS>+CzDm6xpl+Dqr;U
z+AthaOztlGFAyz@@<19)&}wThgp!7K8>PGp2wEndem;~c`6R1hN%?_Xh{R=iEOgHz
zk!g;92UzOjkj-Wr=d#@Wv0tv66Ri19QdCWaI;_9-)CO`t`9ux=n1V%kKi9d>*iut`_Hi?P5QT2Pp8%^-Hq5v=4B(!IlK$_v;8VJe4*%_lkUjqslX#Q6F^bcP?NqZK5g3T4{
z<|!sKoSjl@kua*5sC7y|tEB?}Gnu_C5cYp1VA*=Y)QHmQ(gy1{NfSKoi
z?1Aml?%?WJlwxrTZA(6|v4@d>?v|EVTbKKIFUT4y`L(oEqly?kwp90c`qH1JpJ8z*
z|4TX~%7#0`j*(Elu&~w#1I1Ej1eZxQq>CBs<&r3Dv?p%hV)N}11_HJb`--78X>-j_
zwPuD*R5A@pwprRQu#Ptgmzmsa4@dNW?9nGrf>!nJh$6)eDweD_B*_+_c3)7src@8c
zZpk?{uV9n0TUz~Ml!R80%*%%*$w}$%*IT;J#TxUTzvikito~60kWf_`XT^M7wXu3|)PjfH$PG!!QkOd&F!pm3T;VM46#T;NfeBWBU6@rp
zert^EMhAhPr@n1&J;X*yuS!VRo?u9auJQgp6QQ
zcEQ}=SHP)8<2F<;k@sz0*gjtI6GpcDq7)t6V!;Bl@V^pWrm@$bfQdoBn4D;^W(`uD
z8nk1@jMN_xn|wMOlZatY>H1l+^2Myp*o*AU>o=MH?-hftux3ssGwS+(p|3PER2%TE
zmi-1txL!Jf05pk1jOWrD4dI@v9OjAj
zoN&v~R6e6xPzpp`Od1n^KK3?FJ(p9LLb1~;-0)qT4z+kEZC0CyVe&s~5uhP)6TPkZ
zb!b?YiL8gV(Ij$G$s1i>;k@u(J{BL>nvphRNTO_wufs56K9?^C9=E4+Ar%l-Q0u_s
z>(4Rv_V;?=?>dF%gi&@(ViK~;GKmUN9Y`}<1|w#t=g!3rI73>0RPst5L9(kwj!!F5
zZ1_YeZ@gG@<;baEva<$ZWKwDX8N9lodd!icYyj1M+pVD)g1NpJ!SoKT$z8NJ)BO0__ND`fY!Z1t>T$Y^jHWJS-A$kg%VS32W
z7Cm(zY=)(lp}`u?v!BQ#&;3x;`9^P#V6(!JLkqB^YoG!8GL$Dsppqo*-=dO0yv@lF
z6FH?5c9eWTx|gh*kelY2`q3uMe?9l*cYb$ROKX^;c;8kXSX!hXoB}O1Yp(}h56w;v
zCKW(`W^&KqaHAQu2AA}a%Q+9dUUPx=XF+OYOpr@t_G6O){*u1WQd5TOTG0*|gu$0;
zR|`3Ds|a%45KEcOIJYc0LAH@>yzo!o*Gvu2dY#9?H9<3!A`
z?d|p66-+lMOkyvQ-f05*FL%+~Gvu}Si_NM}FMhIk8;?vUJ>M4fT6Oxn>u(|5kBYv3
z*NsyMA~h<}IboC=M*2#(O<=Tv6ZnV9d?dvJcl`bl20@ca_C*ZE3ZkSM%G5wWS3y*yn19
zkl}Ha4mNx{Cz}4k6V4?I^N8ea{c@Clq=+_Q>+z#r2e*`=OSB&<;h(8@b3b=Rl*3{a
z_Zhm9Ib@EEs!_}5|Dqb!38L{D;6%p5L40S~qZ~
z87=3CCdD>5J5wxyb1%B%K50Gys<=r(l%D_ou}5Nmxx>$VG;-dOEygvee;AFd7JK3C7@42|S78B_YU~B(+>{7ZxOHnZ#Uwmj@7FLts{a
z`}I~hex;1gSc%(wES&j}%Ld0;G&H|7vS4*QX-xujkWhWp`gAv(>Tz*M=BRMxxd3xO
zjKBRtAky~z5x>LQFrduP7_kt>R-qU|FTc2`y`DQoa=$t6f>0pJERKI%N?i&I4&_@o
z$I2DgSg%gzO!59!P0Pv#IznsYf85(TPy%=iL-$~q@ps$51E@S9iA
zTOIDFN$#IM5(`VGELmIfIEWtblkJh=hHP0_9@&E@c%!n9eLpRjP+%n0f5&8R{8Yt$
zwGF9!@fgrqu!Y8fMBz~Ex%KRkat)jS*@&ozY>?+cmC1x60t-Bi!h)UH-AYSZzQ^!o
zAg|oi^QVw4UMNh6BT!DkjoNIOWppQg-@H7h%mMJR;TIIWzIIAB3xBrU6tjUmyOzeb
z3yCPnRUf6v>+VU)x4Nn*l#pNRAI;TviD;xt-rXd9T_*(k-8{-)-&
z8(;o9*T8E`Vx!!6H$MItj}b9O@n<{NA`nxaELxR%@wFi5?o3rA1huJbc;zxfLeq
zg3&c4^uF9odeF1L{Eb`HZm9=L?m8tDLq>BH_48wBg>qLLSxqZbmh~jPgr5`{XfZ^9$=3M|hz*V_XyBj@`8%*uvlYs=inj
z;A6@J>iCGU=?d@Ar(++%m`auVm~9yTlPCu%5EOcCk>Az5Hpcb;xnA8xgB#fGs4i7#Bg4?#Z
zlT8ECKCfYIr2M#_Y>#Lh)wWL&>#=ej%7e^E9WuQOpC{y&x$Y1@!7gy+e_qz7QHeWDe=2D!@Kx|0@LHz~~;tTns_zb9mnH>%-_;v9~fnF3v$
zo0q#&=nOxm_}XlT+$Jz#Fb7g7Zb^CC%|BX|JK5T3Bj$a$_P5C6LGYDl*8tx03k`}u-JQ5XHLy1-?w;2GVtf^;_628-xmRDy2C~_KRgLbfBsrru+e+!!m%p=3TI8bVI1Z~gYEVzg=F3w&Q0b`
    hH~;5Xsv<||Db(eGwI*Oi=YNf6*PwHF}IjOOGTZs@zstuvH!Mesqwd50$E4|3eu;3nm{W`g`D%w>_5aUmQ>dey=4k2H0Kux&
zg~f198Geg6Ag@=C^m!Gh8JyW73gc$LeS*mya{%d@V2!f%I(t=`(ssw^K>(!2h%RB0
zR>FLX+K--m-^L4%!@Y=r*1&Q)f4SV3?D%9o?V|_Q+zJNS&i8c-n)M;M0{z(dfs3$7
zJK2bNZ+VZ#J7HT{ta8^i>XA2$H6Rk)rY88J6nh=Y&%6mwqSKlLZ+GY_gn83tII*kF
z{d5hB+&N7OaD!$)g_INV*lbn|!8lIbn@~}3N$6WPEEr{~VyE7|xfA&9=)onGZ
z7?h$mQQnDQbm0Mq*RK@&lPX)4ra9IBo?_s$IYR)saT}5b3W9(gV*r1j}6qQr$Prt(*K=KLBJ+XPciF|LsKd
zL6TBk8LV1NmZ+!R?w0PBdxvf^S%AMnHJNUt&-9=KOgqj5RcmT*e^N}^J|dQ0FQu$P
z(L*`0RV*Zn>;R!&pkz*Npt92ZMYt!yPWNWvXAdsvxiMnz3OiP_Rx`t3R{BF0^lYRk
zLruPql4mH@(U;I%O`|{*)F0*!S&``kQiYomj~I^X>E4lLvnn%ZhqfOaqJGOxc@y*o7B-VI493=_?70(oyrvQ6`OpXyPRQn|3oH8+>?60o8xe%_^(f8p;=SQ(
z>ay=@ckyexZjk-fcA_C)@f_w~+FJV;_#x0|{|jKIjwbAIvzv&8CdYwibJxflQJs=;
z!?}>eaGSYNf3i%;{$^I7?NGBR?Q-A$6cgrPp{vWefY<3yc7aI2F|uVP(R1*6-{Qim
z^D3eg;Gozk8Y?TcbY63tOf&;Hk7zyCdw2Rm@l*YFz
z#{)XogBj3+kKd^Y*u5S+RelZjYxItZCB{|9vKoQ8fBQs}H+^f8gB_kVWmVc5B5(?a
z!kL}kSDRo&PN+5@j1>)ALSl1Tf)`A@71>Y`tyd6X{uK{(5h<^bIv{bg;gXDU7Lep^
z6i~EP$`nUs{CyVxBu~LUml19pKig)Vdn#C`F&>;R^9*q6rDBxXi{_;%KDi5toPUYd
zPX78)e?XtaWWKsS0A4k+=E8B?pdM$QHI9$V&2q~Cypw_k@mkfXU+`jDZ%e486b3g*
zjI7@vxGV>48*rD{HHl*`?_rK^jxOVZ6Mmv^>jxl>GTPrkIUrLAPZO1a+mvxcMH{T;
zo5}yPlFNY)73CZ}g4Nhm5+2Hw(Um;=Y%FV*f9dUYZ_0!IZ$kSuC2fzh}CLcpK7@|e839DAOJkUz-*FNf$%6Hmj9xN}<0=N`fmN!HgNKDY5M
ze{**oC#!4kn~>=)17q;Dr&Rm=QN`t`hVVx@MQH7tkW0|0dR9P7f6(iHN3LrqGWS3&
zf|NnP3$r2`>Xdx%NbN$7X1G2KclRVv8yJTeEMOVKCWt>y#hmX!|4VJpo+y3;io;*H
z#Vl!9=J;G?R9D;87B|HjU-Ey!GAr*ee=>I|^@;2HAz`T)$=5mUoWrHZO42EsgeHjUUclVtt;u8dI(t8(rU4i%>H@6e16fOvmbgo
zN#)R1DJA5O>abvhys5e`VXMFRYo5o(SKSi#0rSgy~5d@1wvic7RyAE0^tOe}WWw(LCyG
z|D$8r1p1^P6R-NwKpRxy>cpY!K+T>-0xE%%Oy#+kI^TiU?3QC~`?$hj(BsyBzw!vQ
z7AFWQM3#f*?>0M_@t56u6AejL!HortUB$p0*BxPqzNR+L>q5q7N_h)DT)@KVku
zpho)7LIxu1G_>R0Eo;|(e?rvpWNgwH7FdIYRg^+l^B7JjLWbg=dg9ooL$h00Y~aPx
zQ?LS!G3R?`wGzWUDbAlcl!!eYEN|1`zPu5^ws1=?@3t=GLAKBHec>x@GSGH&;vvX)
zGZYW!LN9d`3~sOIqW3ttEj;p1Vne)oTiCpwdWQ`x^+J2kHzq1Ey2{I6KJG!mt#`r}
zhm%RKzr`E{c{wT_RpvYNDrPEa#xM_euj@jb@RM=Vns+tKe+a%GG%UaFZaq}66IyNj
zhGFwPSm$#Fy7HKqQJKa@GF{Swe}-TsDD><4z_q7r2~utZ@?#s*(g$_Lzunoe&>F}9
z2@x6IT&LQKYHf#SoUVqzCazYx58v_UYjn=?#8S#53LdZrlTy@Lf;ty%gNMpO3Pc%z
z)97YkLnWT`e}Rx)<1d&<`J^rM8pUk=?N~=9fNmrAEo4caZ$aQiE~1e~A8O$&n;5PU
zlU4@(`OwVOU(cgjk^D|{8a>$~3Qx?lugIpd$?7le@%aESOfT-IF#grZ#mFphaik@(
zDdppZI|2%P0S*43`f0oJK^tyO?p6?p%iYDvlx=mCe
zrb{R6dGGl|ByISS6Bpi0b=H-FYDQb@a+UWwqNNU}XAvdStA$P7eZl|yAV~rxor<6K
zx;4)p9p;VK+d+!Om3>g&J{o#*K9Q10bd<8?PS`qOt$4ab7Sv0rn6%M;JC!OFRXXkN
zXBmA5f2pe8iV*Qma2w$vrsK#m{lgp;M!yJErgA)+(R~`}c_*8}<_(12%CD0d8Wl9U
z5b5AD3$_u&=|5!~kPMop|VRCh}zOq-wBaK#}{Vti~
z)?P=nlvZBIB~&xQ@yS&v>rOAg5N7yXxmJp)#*7GFhqEr+ul(&Vsi9+Y0(>FvD1lUW
zf0=&jCR4$QlVcr%#KD**GRAGic(I6;<@Mmw+z(2-1{cTPO`plo_bO?3e?r$I>l_BT
zNa{ogjg+a9>flyzcV({swyRS*T-hv0RK~<6?t?<3{bHGX_{EoHkO;*59-<5rIO{zd
zrwC;`>~DI;g73p7Gao?2I>O_I%TZ-he|qg=pbLO`w4M6=j0i{_YrQlsd2jNC@CfLE
zUBQ9(yt51^a29;xXybs+VGeDb8FH_Br5LI0uH}wcb
z)2Vw+EU~Z|{J+1T)HaMOuLO2_f24wtoy_x{6-1mzq*>$g=bszG$4z+zK^`3KVUBtvh)-R3Vq;$du}lSD_+*Z+`M!wOOB>R2(M_DNIEE7
zzT)z%m&d66GuGaL^sUq9d;Ti$74VppOfef?oL0>Q0rfBwyPY8agi
zl~42T33zpt*c-?0wRH|1GJgmal_Otr$IM53*OVqwz!NKjYcRSHd4E`JqL^Uxicxz
ze*m$Boz+UT|K~@{Sq~b%e>cf&V?&Px^gXw3mVcpGcb3Kp{(7Gfd_G7aW1pY-(XyPN
z*f;aLZvQw~O7D5dAu8rc1{@D~+)Jz#!?93B9PvN*m^ImJxAoFd>J73G(e>LQN)c9(D?SON8
zJEcfW#=-&Lhw$+Q$C!fF78RFl@|M
z3&E1c3K(7&|a8^;Ty&3i9I=yp9WRqP;uw`MFfVVBJOBXTS7G
z0jg&z@wvvx`^xe%{7hDS{eOb)a&5kE#u37`t1#_spg}iUFExRHUqbn8P}6&s49qmx
z>_Szk-59y?Yp%9<0
zSPD9>bUvveNNOS7H?4lnko(|V<@J{C0qaL=t&s(e4ESz}c!V4!^Qx3@&9%U@T_yP3
zF7G~pgldft_5|Sy)(c}@FLKSXEU>9MCrk&}e^M#Cvb!%2v9rcwgai|>aku^51~~!A
zXCHU-N@Me3#aEui;0#<~4H?66+^jpVN%wi-JI#iC@_K?I-mp}1LTHATgT(uN@|G|%
zlzLI$tDxMQYE21>C{J{5y!uvi&AFh(ym+CE#;zF&Sl(T3%C)No6cllsN#HDdH=;#E
zf0R0rUW~@*Cpro>j`1rkU#EA8^aI!n!~15}p3rS2a$!l<9n6PO8GQFIlFI$nr)RHpNDWjx1n05+Q)OIE-SQd5{0h{d&
z)l{u9tC^2jA-U6aM0j4uRLI45H3J)Ze^}AXg^S0u)rri$&RsApk_Vu1)f00zeP~J8
zsL;Qgy5%6L&-t@C
zdN@~QSnOKV7jic60*;=cY>~-WB;XDWOcUKmrWLQQKRg@@XNSb7?qW<-0Yj%)f4}u^
zD;{lejZn@Q5E=!RoUhZY%Td^jp9^uM9@34$1DtCM0W~INW*@z5nu$kKl8b@;P%8}1
zq*M&rXIe_*Zvu53_g)LnG2_eWC?X|m+UTUk%eCY@A9?~4dJ>m`%yt2)E$Cc@l*KXSL7Nbgk7(u
zAe7zlo-o?z#8v_Y8gql5r>9=6i$9q(QU_~z3ZkcP3O>*mmP3gu9vy84=s5QzTN6i;
z9gUM^p2)vblu(WW7yo5+br(dIDU?^f%FQ-FO_zGrG2|O)52=o?bf|lpe-KT5D5yAk
z_(1Zm9bP2hG~Lmi*o$=7{k~)*?v%=$KdK@VQLx3FpKQk)C38i_&oiAfGg>&A
z*;;Xz%z{r<`Hbboc%?7&aN#7jpq=;E5pRq_-5G7+0!uJo?oEES+c+hw8f1(AEHa{_
zj#HhJMfA0gYmPoQ3=!Q`mepf4v2MNLZK?ds{%JEwp4=7kF}Lb)S;;Yc!Obn;<}8Dp
zIH}G2*$&SQAB}JTe{$pX0Sb{ps{H12J2xBiU@-fSWVVikt3S+I?>F0Un(&CnFn4QwiONclU|wHK8bC@cm7q@U&(vMGW@?e{dR;v(?Pe_7BtA!-iwz
z5a7QroiRl~pA8DWH=4HFo}1klHrYVU8zHw1aM{xC^~(+EqdVnBUXwA^`K}E_j>^c{
zOL^$=N4?t}#s`RA68rjdLzxN{!T=+SfQIf@&AL_7l&fZ^#M7%R`ea6fixHN$d-n)*
zMlEfX8^ujwf7agevyamdkN(UKA{l8VbwdK--U}`}=+#vbTkM6QH!Y`Ys%EhE$87dZ
zge8xY{=+meAn
z)8$6k7h)_E9M7jBCTbIXDP>+re(JYm8q(rn4T>FPProI8jFMmY@jH!_tMMmZgOrcI(7jsQZ0=k*k-X>jo(B#j?ydHr916RrH|DD1`D_tv*EC0z`0v
z^XaMdt>pN^K?dO0YmM#1WV}p5VXc2zGrHIx+}bVJDb6;+{<(nLgad!SUaRODj
ze+J34Qg<3(RmsaJRn^c%oyVkwTSOoc@<6+Dkk_8GSDZ(!J(IN>
z_s`W(cZ}J3}d+FLYjs~!Dj$3z9cF~Xt)0eZ^bMJ`lY)zKvyKbDZ;^R
zPGkTafThncB>>a&LXC($7HBE5BcASPf7#4>O|X5^aE<&B%wlH4*nh1Dld_|uMEUK_
z?AEpRkakzRv~Z&YDmGlB{@2AtyiFm;>f2h=3F6ZyMSThLLS$H71e^J=!DA@+nJN;8
z`{V&`%FWQF?hg@dU}6?qp5VF+hYGYyv1mCoy%-IG?e=fF<8e7H(B+wc`itVee-m#U
z)0L8-e}rP=0(%--ZUF&SbX;$$5O}Tsf^6h(=dwa-3Db-Ke<}q*AO!WLsNM+=%
zo%Ffrq~I9
zqtgPh1kf_X>1lE1HDr}!$Z2m=gAXNl{#dSFNggU%q%K~_9n_zv?~C+Ke_OV(H}3Gsl+y%do_u;_YP>G!p
zzIN;e?$ZjT0NYcp%^Oo0YcV={bBC`Ehapp6OwB>k32r*S7i8(oYD`4Q6;5lgP}Pr$
ze}{lkRO3c0LiBAM)a|b3f2x(egJFlIi69hijznxrGEW~lPOR*r8&`fbed`-;VJekA
z^Ak{3>G-XMH$VTnvQT@>^-%dab)YR*0%A>fS;1eY1x4Q#3?A0R;ZpXh6ViojeBk_o
zOq5VT*ys|OjY9f35Pc4vdJeI}erln%Fhyg;I@^b6y`+wG;LAHFe{aX2J6Vw_REu4T
z44{H_@^GHz88<>_U@P=s8f$2mWS?M5Ec=ZU4?|;D2ep(ThYmEj&lKniRGP;B;wEDedaJ^DZ9n8
z!t{cW3lmVzLftBQr<)tA1k7zcBEmr@Wva?V7`|AWP}C)`ME%Mp#`{=4TE+!fRe=M30@*QH@>g`(2W2TPivPqJ;*J&@29gKy#+
z)-)uY;?3-xZr=T4ZaD5-MW@@idnv0>{!IwOo(uV^e;zgeppGjhze^+pTXVqQmONfO
zU2$nUcvt?A3Ltj&D#K{{n}v&=iklr^)e{n3}+K^lzV8hs<2~1)afSGqRhi`m#GNFlYVfd86kpUIj@5zFr?&bjhR`R=Z
z!IO++d^yqYV%pBB_Jz#JqpT8bn=%p2D;AUW{BEgT;{Y|<)??0^4~_zHZz)^Op9+B5
z@CEauxIqCzRu3nLGg3E4YWC8o0?IDL?i}b1e-)6um7m%cj(CB4a?RhYlm2Rnr_vWv
z*NVI)Cv=f^kAQYfl>Ulz++c0Yha?27ip}Bf7O*W4{AF&><)KOgleqCy&TK_F6OC9y
zx)mZFmlmw9ZTf^7exA_-Ok^?WrUO$ot9Kk?A8A_FlJ#K=b@@(NXAl!)TgFDOo<@7z
ze-aytK|hdPc0vf2#zldKRO$x+)H^6adI*`y(&3eLcz6|@KDB4^=_O~57$J<}_vYHMPhquv4MpggJ3Ef6{i$`*`_WNy>_$cW*xiDRU}*P{Kmh1^9?t
z5p@TI;W@8W)VV55Lo%6ABWBj+N-_F)H8rv5q^^nKe4pe{0`R
zVJNBJkGrKf9#Z4B>~f}VTcvRNrB<)K#g?3YYUZbLG;rLY0r(wOdGIb`e}OdJL{yATc|
zGO#B147<*Lg`Z^WSsuoGe@xbKo!Pj>F>(9avY_cDUGVj^y-leJ&|IBcyfOB!4+blx
z5nxVz#H*4|m|g`K`8j$mHgM7=!0Zys&*GT)tb;1_zMnz8dfJ!Cf5jHf*V6K{PcKvG
zo%oy+2}~OVUt!=O+Z5J*wlewm53mkKqY4AU16|t#H97H{mU8?hI&`we(H$~>eQCSS
zjcv>%o;hgYgc)#GDg5EX*&E<&Yn9ydWKeZ-W0MLm3;O;MmLk_nU@-7k!hCA3le;rY
zEi%lWUhrSm=;22Ye-$V`L{IJG{M^2(h4D>kH~`a2CpqY_vIRbAe#z3Z*$1yxlRK4*
zVizaAvGyjtM#pivSL>V&97HJUJ`2CIkgeBrSxssQeTVfgk?
z??008aB6QaPzkb+*w2a|2&jt<`awR^jm-eL(mFV?uDJaOe={Y;($^r975Q0f%24T8
zOvZQeU$pyp!?loU>RR-draQB})H}Y<)!e}m*pachF8`t!}Ti2=j-yx5I(BL
zaD&xQouVpQp9=PC>$Q2Kps)AWD0pP}%(8?KMpf1qk<2eySGfC(iJ>hmy+;Ioc*@BA
zLv{)w=%|pJ@85Y)Qto9Iy2@Vn}riaM4qK*NPe87>Ise;k_F9Oi;}G)4v5BH%nb_>?)nj))?_
zC-@S@aF5%JfNodR3X6L;I%f;8n7X6c0opRKf1-=v!G)($6X!wg7^EJ<=QT9ny7El{
zcS%zhKg7k-5hkW(I59DT<{$h8?7|djXL7=0;vIZrSzxWdr^_s73vs<%&zOGhvz!gw
zLSxctPAr?I(T_x>DToq>Gc
zf9!ps&ts{CqzmU(v3Lr>{cGizAtswIn3$e(NU?t?87D<0{9(Go(^||n|ALBlyo!v0
zhve6;LX5y9*KQ*$+zc*>JZWz#kti8`@L{L}mAxy{R0BNDdteLw7ewhk$epFvJWnq;wmE;foTX7*qUP7XFr
zkhQD5nKOWsor{A5i-AGf8D!=HhC*b_e_TL908Ty^8-Nzb(FNpS4sr&-MFg=J0P-LR
z$QjPJ1ekjQR6s6f+MbReP5`snZ=eo^xv-g=!QgBV#2O3%vA|WNp^l!;U~3zfUpBbe
z*nV01r7gt{P%^WygSx})zyLFdB|wQ?g&m*@b%ztd0A?rzU=Fe|v$q03tpM5}e?5SX
zmaL{0KweW#M_r4B9d205)zJ~^{6DcsYiaAqKL^N2s%pyuKrf#I6$An^FD
zp9562;rw5I;DW!@Rb;g#we{6yIf1`&0B{1_K+Z7mFSq{{8zVe2z@O0Is#eZWhu;nW
z%r-7AjzU17ySqEPwJXep9qMe&f9`1iJ3eh2Fbv=hb+!Y*pUxnA(C@;yLM-9sbg=>b
zW8l~60F=QNAP5ZfOGys;57Pl&O1Ki7?(*-z;DvDcWz+s2aR3Yi`VTiYX0YG7%IfOM
z00%QL#03N~gIK@?UCdluVF1(LH2522$@EWvAb_;1v-2+>D*s|R|J%*Ke?pgn!ZS0p
z_wq4w|L+qsgSf(8|5cm+Oxpqqfq`KzuzxIq09Igo(68uWzs?K{`At)iR8^Fd)zW5D
zhW8$XO$7?C4usvs!{xX1FMEK^_1n{Qav>Wg(W*PzMKi)G(}H<&y!!>*NA;
z_5}VP4Q>a4xl7?q9f3L!u#+oDQRaVi;3TX+G;5FxfCB(>0)RX$
zY=FO9|L&Y$WX@k?cu9P`9HEW?D>Hi-$Omi%f(SU^>%3%og)
z|NqR~fBBMgwYOI_a{w{_?}hzuX)_0~z32ZT|6g*tpkHw_t3sU}%6IUe~SDM{U1}35Nmr7fQ^%for9Z?>mQEJuW7S~w>5kSA6z=G}CWkS%?MH5)82haPjg1%$%LgJh3?7
z9m2)S3-IEEH@+pv<97=Ifb0;c3tR=@=<4DFu!1^c{W@koJ^)ba7wI3w4**L4L4p9F
z?7tB|2LLGd2XO&_3V#qc0I2u}@dAJ^{vfyqpFz5!#^VM?7@FQ
zcoA*?g7B>DfBu5-9PIz{_Y0aim|OmF$p;g8kw**`L$>cX3W^V?w`AY|G
z$K|hl;3nYf`|k?E7Z}j>F9=V@{SV}VC*$!KgeT+qe;0%&^ZGaZ&#_bgbu0Z|vK)Wr
z@PDrB-&o7V8EOa81zW;@to$VcpG{}5hXDtCLvq6D@Ylb-8U2R?hCj>wFKsC)sD~FD
zHy417ix)n)96SQ>`Qh;S4^4}I?#)@^caGP?eABc
zlPbqAEHxxH&3sjS$?;?*Kjyq-*9pv2S@19Gpk>)r%6X?>%rlkQ65m1v(D}G0E!ft1
z9nJOW;qh4ttzlWqavrLC;S6UZk+VI4$HF7Be|GYPiFQPnhijtPD6<~PMb>H4ccCO7
zmCzi!I|;ucMiq+RMlyfR#)l+}nNa>UtPS0~hx5o%{l#KWz7Tc=JpN25LwqrGDYb2io!zt45
ze+AsvgoovHm%g+E++Oc4`RK8(F+i0x`~BY1Mv77NFDKru#&}ZJZv7-(Z+Dit1J{ZdXFe)+Shh$s3VrCo$ZBJL(W3FC9}|9)+tdS<+vzRY
z0+QTU!fF0uQgF`p0S1b~?bwS~3$k2ce+zblOx+Si@4J7{3kupE4yDe@4P4PbrVXSH
zzt7C$d4{62n=kS;rF?e&(vG5Snlm@bVTQjrhFDf;v$ovIa=bmEIgQ;O|5YWU2a43I
z15F`Pg>)e{B>gS5cokl;N#;FWmpy3M>fK~p{wqnb!g*t<&3Y@XOc$bR%rg$HfCl3WM3dCj=^6J#rVt{5Z-L+RNdF5$Sd?;
z@%8a{Nj1FKtoB@Wtx$E=`P6lWf9H^dFqJv9*akdZkB&jY=WgG_$d53N8=DdcwYTa#
zCKon4d#OCOfIXq0XXVl?75~}UW?bI}vZ&8=b2Y|Aulvl#7txbiVwQ{!JBQ$S!PYM>
zH#O=E@RL*N183?z#9NKFuTxp&vSRRfd!SY`?Zx`T$`hjU=1)N{^qcgre-b#6sCd3~
z{1}XWnfk@Bex8lh#6OWZob!|
zA=6D6TX@5yya(|G=^Oup3!LdnDwef;d2sn~;%Y>71Rj%uZ(-o9iBEIH`j0VR7p;e+
zDpQpv+8}5d;w(;UW@>Nde>w-Q>U1)FP}pkS&&m{$H+TTq%y?6RPa~a*nCx?j@5$@b
ztjuL9D}YbYYf4KV7fsED(#b1S#(&g!u-uqb7GgFs`Phmw@6(rT9840MF|fzZFx2n9
zzPG0rgo=w_9}$@E(^g$(*u{x56Z?-xcG1$?%PsOrM=TS*U!kC@IJ0R2V1dx5i@pOzmf6-8aP)*EVswV@cZK^%Ie6_9Dh8
z1%@GiXYQS%`co*we;bpon?Mj{C;W55PKeRdV+-KUIXIK4942U4gx!b0@r8@?Lw)dv
z>a!M#SsY->YQNWnfB=5(R#?nt`6b(N<$az!!5yxhR{?EpuJ^Wl)GBFa(X%O%^xF&%
z{EfN~8b0hYqG%z$Fp7=K^uX>J#gB4Fsu9|ecT;a)uAHIoe?mt>dY*>_(mA6?y}xr@
zJ~pbpX_qHR1Sn$TpBF`;0l=Cgi@Mq;cad0TGi0?KafXgvJZA;MZ=;PK&exhz1163b
zqK+rAP1P~EtIFTXn-T0h^jS={veq-1-r!+Vf4E(1-^GIe6T^$zzAhZNBJ4_>1u(iA
zn!*rMV;1i{fAx>pzZoSko{C~R=ns`eXbQD_twBr9Hw}omBa34>#oeI91{SuN*vn((PpBI~KkiyH9Q=y{G>sKvP-2%}~CRda|QD
z(BR}7>8bSo^WB=z8_TFTOTkKY1j79~UxE{hAk94+e`3QxFS02$=PT6QZ(W5h3A`53
zf=31&jniJ37C*}xz)}cN2tP)+Am3W?l-McTG6&y)ZbjC={lMz2&LF?Zh{`D$_i8oS
z@mF>edbySL(7omnV_4OISuV$N*N_-j7ws0LJ+K!
z86ohQHzh19B~7qSTuB5+tZhx=q-@D+rv!7mTY_fhEG(&_zyJJ9t8+_Xib}_XUL-|5
z4t;1fd~jgttn%K(mvHV2zD-@?Pb6!PouTEXf0}J644B0aDFb!i!)GNO+bb|S2IR`t
zC8=Xq2-C*bd*|Vyv%LgYlOcV+Tt^G@14>_GLhis@;j|Yo47D(#$F3q!IFCYfxmJ9J
zY*%P#E1r2E8!}D{mTZ^Iijs*n&ix1tc`;^cpxzA)9aO_nyAMTc~Up;X`_7@rs`mAH`f0^$8sfI>+6Qv+eY6
zh08V*uKOyZ?J|U$%`5+X2*e`hPa>0%&FU7Rj9t{c}#aHMSPSr|l;tBfxF
zY@>05r8CsX7(nd9UF-|eYwP4YUQKkoQU&-7cMNh?no8Vd`m5#1D`dQ-)@#%FVzRnw
z8WkNMBbn+I-=$_fu%b69#&hEnu;NwKV-MUrH^8yjGpe5+$x`$v`z9!huH~RrxvZ`OWsq%;&DXE1u~Lid=PP
z9uWmPYHMiqI+mWA-6&oIpu9bgz&fdZZ{HJxQjszj(lR!N5f*iGC#ZysM2z}b)hMk(
zeg_8;t6P1*yKZ4X9vrU*tJGgVe{b}b-<2FNmn{fg3jXx@D((z#c+*mN(v}pTyYVv?
z4uRB);(pddwnK&2QKC>QbQ66*+;>~)gE+KHZkOb3_^T5o9;_FJ3nbSm%tg$IIBTQs
z?mC|S(~#%mTw?TQj|gNsjNZj%8@1KAyCxGywX`=#^>%Tkhg|6PcF2ajf3lQ~$o*I(
zt<0wKq;Rq$&6E4d(I@-~IVz*0H_f3G=Rhg!oVe8(pCZf2PksCwZqfAW*Bl%l^)Qdq
zb}u4xUD@tNl<_k9)$toKbgX|=ioYm{7`_av!kGBft_t)q5?mI2*xIu)>`R0A
z1uZ+?K{x6-#5pfwTUGI(e_W%5;nL@)%$hc#?hB1D1Hv-RyEo2s%O)Z^Z!$Tq3BN}~
zW$!PR*c~(;TSDghHf>q@q_kM#xirT>4
z#4}~&2!c%t*+yP4tmmfuj)aZFA4Q2@n1apMOY$)q!N)-Niv=B*pIlOYFo1Wh1`il5
z=Q`I%*2L^ABV4BDIPoAEi|6OV3YX@7*(It}sC@QbK1=`Te`6cbX}A867i|oP(#Ngn
zA)n~9We_Se_>5}7Kce1~$?0u%Z(ejn0=WB#3<*`I7F(36@pDP~WiYKRd`pXt>=Ewk
z3-T+6vvnvE)D5buYoE7n6RVH`_WSYso?k>yNtqrR(y?0y-`_|0zAuZDZ$o~xJ??&-
z{bt+YnEf-hf0_1NVt`KCG$4>^!e#;trOtl;^JES+Bwbk;$15cgv`*$8ECFPLJZ#iH
zBYtiJTEn%TXM2-k{s|goqRF@APO!AFG_d=0Q%>KcUJI40g>&GE;kzfRgcBrcQQxtt
z&G>%ss`uGT2#@ZCF#rhXuM2@>*3>k%&6$_QBjxcS0!h1<(?tEqDBb~8BC)IpUcgsnT{|IXK6N_g=XSlm
zdTya>iec0J&e3aF_NdR5Jx(=P`0|sLg%1wwb?flg<1P1`ZZ;F%ZB7pmSxQ&2E$CX~
zu4N9_fA&A|KlIWON=I(a`ufvO|4ZGlW)1bG;JZRpW-7k(D2Y-%Tih?Xj>l0Gqx6LV
z^a`4O!(gs7tQ{VVa{U=lqaAkN?IU3ekT_@&3!DQkF7LUKrF#o|O?coQAVk@Rp
zpy0mk_7oEFQ|Y>q#W&*U>ou`j=&Xfa}<}cUJ$d(@YbY&K?H)~86`F<#To2FQ9&Pvo
zf%VyY{2<%lTgJP!qAp@W<5nUHmbdwb%|DfJoA1ijq?F?1wO({gAA2pZozH1aia@{)
zdWQ-5&kTkid_88yn_QgywCQo{j??K!HW34;-RK`Zlk1hfQ*9Og^!-UZOxjC)UT$nK
znDY4}<-Vochs)&)1lL6gLwfQwyfu6QIBiY*wajZevid-a16*_u2xPN;p?XXW$i^_L&)i>}s3>&Rc<=Y$0ydSqbX=CyMe?UZyz@}sXwr4v1X1#%e@6+y)pfFnkC
z9b!zqWG@{tX9+c6aE~Zst
z4v!G6e7?Wo!Nhp20$!g=`EG1*vL9hW_xd9QSC^g@AFnhjU=|-j=^6=w3I(Uhr(@IS
z=<`~8b*aNVLm0%a;-d2Id~3wOG`t6yW+$~2Yhj}ox*p#b*R5H-XBhS0W^h)%$FLyf
zbm#cZg|fTV!ESD%X?C!5<%wwijmPIwj!c}L1P
zcyT-pULzKqc1pvDlRzpj(4zR2Kc|SjVno*uXC<-qEi8n(?(nRzKHR-L7afXlHzb`e`Q{ha-xoQ6&w9-}IGrY}u?H9b
zDOFpo)YD~#2wm==!G>qU&
zyuc|6!ro`j-^R5miqJ|QB(xs8k~86r2V?5KFd74-Gxoj9_CoRI#|E=H
z3PlV^H7}D;4D(ayjMBgT0^i|Hb7nbU3bLKGXwOe?y(E#n_hH0I{tJz{yC{|E$$m9x
z#a+>go^1V9t=kxlBm3pqxbah7nV<6{mOwnA6GE8O&q+El>SjGZ$`Rg*?B~|-z|SA*
zIA$(3F7Ko0abA|d!YCQS(JPY@$;&TW5ITADZ)4v6Tlu&wF=6(eOUboz6SnIVY0?pP
zwD(bAcsu47aM;f#Ln7JE4(my?7g5K>V^3aU+zVzMqt;hRW^+tKtwo7tPl;7eG$=D9
zlCl!WM&`nLP`Hjf!*~2l45d0@Hr;<|&w0`msm2CqpAU@yh?Qr7kPwsyAG1c-9oIpMv}V&$sOU2kkkBgyKcYSUqzJA*!p?_8f95QyYf^tu~W5b>&~^Ecq3zoTd%3u9WNW*>Gxic
zCWV>~FkRFR7p^rtbKo+{Gg6SZhwG3?5p6${O!n880;C;|WX}xx?`rt*vGlc<`g~Uk)6e
zY|f^CDcp6A2HcM`Aae6D79J`UYMvOl1E%PVmzx-t$`H>}-4HQgP^^|((6Y>~0)<$y
ziE~9PX@9XBsHB5%xPp{WakLorHyMkNep&nuX_aPhRjo#9N!OMoe6F}qB3N!qzs4U!
zH7)u9n7PuYcnZOwBZUDnm1rqH_81V$_)%x*LI7=7L@6%YWwn@
zZ)+y2w(#)}uV=R~GU_=chRix`J{Yo_l-6g8If==lvK%!F@GMg*{vFl8
z%$sHOu5>Ot=McLx>d_0x#+da{g$u1ATfLe*O%2kx%Z%o7Rm7$!^{KpiB(vm{)x9gN
zedjB#&1#e|+MIXwi~#-*X`e+*E<1I2zR47q;&Ny&OtXeHRd
zqle@8!RgTFv&F2sVH3o5=ZzMh6ArC1<|*FPPw0(@en=3QoxmrA6!{E-*XxLfxWGAy
z3}t0Yv*u-hYH8;qMjVHY?)i=!XpNA^h$FNcBY9
z=UB^+E&yhrKPm0Q7*~Zaeh%pekJB(SLX?fuwL~ET+*z~dl
zeP#z)KdLA6mF+&I2at0nt+!!-@@b;3O6G{YqOv`adgubpM;yc?Tur34?86e7oY&LU
z3`e}X2VPOqV$+ww#F0lgvueUnnAG&jLYn3boo8*+_09am!ZK)_3?+DH3A5!{Hp8IqH-Dw)g4s5xVuqfl(^`+q%pp>nJ(q(@5+Azh8em+_A~o5`Xe^KH++1Rl^mhlCk!xh>mUJ|qhl
z8G4c&q8GyA;~aQIb}Qsu6edSp{h0BoZdR)f+YB!E8P=BW{0uK{U;XL69<*Q5r*{)K
zlEa_7l+NzmA4Y1??v*EiNsL>oED6oPortQ+>)u1#%T=<;_rFZrj={P?)4JPTBTtM^
zR7YE-Uz;uU=BI=@DFJf3UybI*%xoG_fp;H8WY&~fv~T{`2SoCE2Yj}ZTcs`awkN%r
zI$HP_^x4cs&+;URpYoL06EZu^``$3p;~gaQ_bepRbC5&ZW7GJ+gF>d4!a)l8SpeUH
zz3{hgv~$euTJiej3}%rQS@wL1_-iKWc+i
zcv@ODR!k?6mzjNctf{9|7Sz}Ap461@$5o(I#Jk-i1Bxu;Gs}oz&Uq2cP!a@H>}cr(
zTcQ|w{mVL*)#2L!D`xsFox93B)K~UBYbCSm;=}ElvwoQgZ&m>
zVz!r$45Df7EnA8S-GNq{8mT>1=;#20M2~fo_~%0i+yV10&lVM!g1*|*DwVe7s{V4s3)0Mi
zE;(V8n+hRNx_wLx^Xj6g7o&J?p@+I=3)H4#iE4!(1z-}O6SE6d*_*JCX9qXtj-_!F
ztLnaWm#aoid?K$L>QEji5M%(1pNt7A50B39q5ox_vT;NJUK6*oeA_p@4QET-u}DNpx^#gzjIHZGG|smFuvS<
zp`i_m6sQ8akQf*bwhlwfW;NOaVSR0W5FdAr&)s!XmgdgjNi5G6Lr9+CDgV~(y!Yo6
z>xQ1ScSf}NBMJ=qML)W0{e;mEZ{rc`kqeSLyeJ-kYOqm10^B`+#6D$$XaJPCZqRyS
z#*ck@z4Z6QohLqlkxR^##R|x=I`HF3g^>H9?jZ>Qce%SJfcD2Hb(W#*u|Pn&JejLP
zPYwzHdk&wd<21t&d>o!?~wNT#`*|^lt6m7
z!%7X1AIH0%kk(3}`QV`J@^k2+K`CbcX^3QaWh^|$C+g)j(E0S0cF*(!Z@5TMBP4Oh
z{)V?~SQmVPE}%tuqfF&yY9W8VCMGqj(gro9jv^_51le2J8loe8iImr%gjJs1Ru4OJ
z2;cT^l-=o6D@HVG{I<0B^SV~XU=Z6Ht^_uqE?LdzbtPo>;3TDB^(OwUD78n_YxDOW
zy}2noqv-RS#Q3)gn$m>;a}93c{d9vESvtCL)hV~{V%PCx25$)YFe2YC=AGUo
zJU$OgYGRZH$dMR|q2Yz|j$@w>O)pX(hp43}SWd0*J40wBE4($9A-A%+X?o0*sF%E=
zJ!FUC{bR0>iWa!#)SJ)u)RD(Ja{jsR%6`?lp1#DB+v@YNekklGoHYYbD@1mpXYRf4
z2wB13&p3AcZ6D3=l!zH}jA9B;*pcAxr(*h7yq265YfosQ{w_#4R2UCx-3`dGiP>(w
z>zUSj7#c_vu&JLb|AD2kYYCjbvkWaIy*n+QOk}r53>xWVv~5eQujQRvH~WJOY(1+7
zDzxa$Q@mpFH|P!N_}0^hz@&_$*C5Qnqir#4fUP0W
z{Fb1^oqGu}PpOj>z>SnK?kf43pwfI$HNju>8%MgYE!;pcS$IOV=9sR-
z$V1V@GUn%E0hz&sZWD?eFR=;=z!X%FsoN*fRn=8;qyXDHBzgApK?47RIR>n5{bepz
zXndY`VSohg6(hn`UcLuazYUyjiN*Q|t4&+-hZ;OymSLU%pLvB_25o%3?uUTJg6>76_L^@-w#wXtr1Umx9Qn
zX98oP-unkFn2GEMy4$EvAn5VRb@nNYl-fB;yNf?hYufLZ;$Kw-jZN)T-H`G8Useal
z5X~&62S4bWyBY_(ct~JSf#5^1^*__Kr>cA`SaeTS$H=X>acHM&*{UTEO+Wo!FhZ2^;Cu8EWgq9T(@S0X
zaMJNXqjl44`YM~U1?+p^UCQ=K#i_{MgjFq@e$mGw@1lf+ICE~eAH+V2T_)7?MI7Cg
z3d#F@)2J4bCiUokEtZi!>oToI3H7;39`Q;;FR=(8@f#{m%pIF+IB8rcMkM6zK_3sX
zRkroGMOj~R3T1`XMU>J^QWb}ay)yQe17sDWS$=)dz$XcC0L&s#-Y27JnD2G8(-UXg
zw_)h9HMn|YccQuD^wVTABl*B8Sv73~O3YS59+
zqz~|;I;7X?@1NxsR6YGi2(U(p(gLUSZ<+VO#3w}d&#;9G(Lb<%>xeA;T`ERkNkR6=
zq7925aC&g10!p63i>b%?W0SRcrud9ANBzWnk-|a)a1q^;1~lsKmAF^MSP;OUD@nlL
zlY_h27s)MLol7+=8<5Au3Ym#JrLCd{4lU2fhl)IgQX7PY&8nr$c#I+xrkuQy^fkFS
zAl&?ob<9%AgTp2BW=^Cv#T;2)s-sB-cckj(^M8fO0ML;Gh;RfS->n)oEe1K_g$8{c
z3IyhMSr?Vjj!0YBcP|hX?G>7^Y#FVdC
z=t9i(ISy?4X(<&Kn@POY9BhkqR2JH=SYlM*seS&{p%pL~)HIXs9BQpJNP_9et87V*
ziMAb+1om2T@e=cU+b^3xL?U1Zt-rw
z#J4m2VqA=;`z^k&lw6vL4m1pWcy=(gYs5>Tdc@DPG1i&{&@5Eh?fY>^O)XiT%`BcL
zz*}c@a2na_mXhHzoor{tE4ui$y?fD(9lg<^16glFec;UiyI>-q@6pe~uZ^0`ZCxOg{ps};Ntr~BjcsRD*4gkCCDJ3e52I&<
zP5)xg71B-gJ5|9ItCuitw^Tb7qa?2-1YZSHdI0v7+>;EtDB
z{ih#w7;+R|){R^aF8pZh8XAcA%I4OD@_M3mgAyT1gS*C`D*JU*pISqn
z+M!{G({`q_pu(C^a_vz+EgYVLqF}
z0uHX!(oZGpC^D(T3XJ%4N2?`7KW^t7=Shr3D6AhiCpzHh0L)=&*nJ%eiom2k?pVQY
zwVj~Xmvz#(upx8#THIvY@*n;ygk^?y-GG5_5YrP3$YgN($7iT~)-lmWMIZ&nP(!9C
zPm;NkAxK~mD^Kf9E1PRKK*U?jR1yYn&H`de3wj&UQ8!nKP|<@LgxHz&sDH%{If{aQ
z_qRV-3Cr}W_$mW1_AK%JlER~iG?;|__mfdJO~Rz5`(>t*XT2Yyk#>LYMYb&{;7nf4
z?!ugAR^@gOg6|-@Bbb>q0Uq23ovIT}3fxj=TPwlbTvF=e0yJ?T_%GHNq81XP*8)l80xnC
z!S+R>;yN4FHD-l@KT)%%
zAAwCAB^M9NNn0*crn+WS9vZ%RQ(JZhVU@P6uPrt6wNL}n4NdesOUUU(g=-DhL0w&Bpw|<`P2Fo
zTa8YZH{#;RJ54l)g_{%9ZzldnjD*Dc3aZ&#T9E*KH&xZ2g$)m=HC`-T9iO%dMa2wv
zW9!P6Wie9B9W4%X?VzD>^!)jKB#DRRaCBi>xX0V&79}i=mrv5NcEdX|B<6Q5XT{2c
zH-fkTpMZV9vJ>J@BqcRe6(hV3zU+CEU*B4MA4yfD)C9?dtEvHRG__|ovsfzI9mj~q
zM%32|&yqIvMxl*k$534K#7e}^Pad(GmlEAWHdv@c3sem-;a)K_@Q5gp03~wm3^_Yl
z^;odO0@rZVGCXh4BNm~_D>HnZ$4HXy1!T6yl8saVs`&^0W?P#MdRra_UZf9R=V}*Q
z2;CZEdk*Q_tT4=n2ll@c@kw4YxXn63T{7%g6ik?*HstpTr>LRUol^Hi4g*rdsXMQH
z8veJdAwHDJ_+GFdqhe{zo7w;~Dlxw(2lWOT);cYKm*wSe`(vocbEk1e?G=geo_JNl
zIqMZ5#e%Ps7%DVgYcv~te1-%5#DMcCryfn2OV=-H@63hXb&DO>m%D9zBU5TChm+4B
zin*p*60#BbjJ&I@1n!wencaG6+!Kc5W?~iW&vJ!sG*yku^%z_xBPC8X1g^dcW(av#JhJa@c<>Q9nbnXS#Sf1peN
z!-dCIz-AP;Vyeazxz@%ur@qINQT$75Ai|F3g*ojLd8vT&H@hL*7w*Tu(-LyY8%Avn
z+i3ivbVS<*f6_om>+6)zj=*JI?_H>ks-E`(z|wo=-|kDM;s1)}hh=HBrA*H#;yt#F%~pJ47b9D30rY0Y@HCczfd~DN3DBIUoCm_
zlXVkByH-`Way)6RGNsRT{A}ANa8EUMB}v^rVQfs)i={ThyEN9OExS3@luH
znc(G|N#_!EFRTKux9Nj9GaX^(?1&e5yP!MT%_iA!QVrg9LtHhGV(O8ClL^p(Nz@W(
za~HAuZ?j_vq(kuhuqXX|%dqJnsek$Q-nEOo;
z&L$HpGB4BR897{FiIdvD>izc+kE9FLx9cBCkc=|GmEOjS;HQdgrbOuGU}yuuW|4X4X}|vBF8bkz(pDsp2%b7_X-J8Bestv2%Bo
z{M9bpz9BrZAeSOO$HqWmHgT81Zl9w%6C2zu-;qZhXsMO;vNpbhcW@NJKiUhv<0IK~
zzJv+&oKi=JoxAJRw(VCz^Nq!j_Vhv68>4lI8)GJfCXnu(tRDNoYD2J2B{l6_m*nB2
z`TX46Rw{r0Uy)>g|5C_njr|+fZf~1rm5nGHdA7%6aNmi`P{Xr}>=G%nf{*HXXbDy}
z=ZOLB)M9~XNy46t>mXey%@15(qFJwE`#IZsjx7DMfdb1~v2+E~vY!U#qPq)~$DTP-
zQ$^nW5e6Fvy?IbTjaUmrw%WK&xhFsTRl0EasCo>ontUvF&8ZAIyPFiq!30HNve+h6
zK`xt7Gf@PrMhdS+in#nAU2-fAfVN68UW}UdRpTVcAq%5>NVw$EJ;&3CXG{MzR$*<1z23n=CwL)LeLKxn)OH^X?!N%mLNQJ#MHhbtp+-R
z+*e8}K(X!aCeWWJjIONUXUV?vhQPfV{$K9eKaNl_|B{Jhr(LkIGHmx&;iPy^RB@Hg
z;ciu`uBD74C4LGAxxIUiMnNYI$z@lWI3GR%qO-e8?zQ7ov;^Dz+oCyq2nOe@ty(^e
zQAfJ5E6mOSu*aDUR-qWI{M{2hKHp3D32<5y6Iu|v)i{hW=7z8u{Z#j8rkuA43b$c~
zp*J}+>bglj!unn~w;?@d^weeZp0`uL4NE2?-Qcp_<=d1$FKZ^n^2m4P;O=xxYz#McIW5UGic-|Fq
zJPC|WX=@K
zQ!g;@?yaI3V~ejti_r0;BMI(nUj_zqug`r&@-sXj+gXw@+Ee5i9EBq;nVXwD{qFFx
zU^0M1+G2ij?M>dPPhw%ILtC{RE&HW^db0Ex1_$c-C>KBFv?lpe>y-srz>hh`ZVo?E
z=J>n()A;VP9YM(G?&kXPjtJ96mF$*+zh;W|si`MRWU{^o~Ca4qN&g-@|U
z1LM_&l`p>^3}4%AOCE8Tj?8LZ=WC#^!Kvn!4&Bw`Nz1q?vITxuyx+E#uR365L8eO))b^$2(%XZdC(PT
zP!9J0^l&yot^dt&wuD}R4w8a!|7X+y6$}Rsl#{u|n-FX;3Y7D|P_-(sJT6eq|3VQC
z!4!EwIXV9WaUuHtq4JMnWPDIVENc9KGdY`@4uKC>LuBix@c5FHkP_{|vh)
zfiI$ha{U)-R0hro`+rN~8YK)E?v
z$O<6Np#hl^r=ygc7F`nFBHI#xf
zDq0Hw4IUfagl`xssWBcRimgp#Um~h2#*;=-cE6
za&`aY-l5!-Jh+SGQ(Bv*&4qB)KWI2gWd#7d;+=2bjnKylm@FDgG|$|;Y!q`@OdOwb
zE}TKGN@b+8Pt0MZOrf(yzMo1qj#ZS*APo&JGq=qQwFs$|JoZ^_^|(fUteoV%Eds0hj@&trV@f>Lj=TP7*P{hM`<ZbNl=2y*k-0s>uW
z9CAfOJ0YFx(x<+23)9|otFoW^a{9m}+;_^kns|l05R%G#p=b$WQdC*&Yr03FRjPi(
zM&jzKuDXH2zPvGk`KII5Af6e$pY24wWrW4Zq2)C~b#S8tM!s1z)zL&Bf=Er$u5kU06_?%6}DH
zARtiLbG~y|aOQRYI4JM)U`XfJJTjh>!MKBQ5ZQVfoG=yW%Fg0&RJGQ=ujIyse=5nT
zKulea+k7n*m}eWneyAc+GY8Z_(VIzDx_+MuEr3<^o_wx;@A8+8)r*k*wb9_vV`IpL
z((6Yu_o0MUJz}(jvB;)OxLAu{%Dk|@Ys^xVCr~Rpq^YlX<8EYW+d}zR($HW!Y_v=`
zgnFZl=jBTsPgobOq}e_1fP}xOL{>>EbJiPwK}5=2i*TW*7iISTAP-I3jZZF@zst$iUoo4*x>?J>a>^W7vd^xDdo}fhTvmkI`E>;hCY;@
z%D#_ajki);SJ2G!i3{BKd2~@n>)&jBsIOPmvcb2*N@M26wKTE!el$btvmAo8rhQR=
z1ISC;RXK%%&R&x|NGRk21H2gdiKlWY)K(Ww|A4)JF)$!a$E#Jry7pP`~`=Vg7WhJQ+Q
zvHvUk5LMIlFYDs22GkXL89tX+=vfmfNgJHmhd#9YlT|o|04s>%nFlaTQ%-7a=jut<
z$y6q6e+V5oe44YkyS@4R3HQ)9*KBB_!tYyo>Q|;M`$+@YR$U*>*E^2G+6x_0f}&2X
zm`!kjPwn;GU>@KY27*xGWCDiGQeMeitAHg44kw9FX9_eF)Ba?RAus7yGZunqc1VWX
zf~;Bm2C3PIqXb@1WsK~t-v+0ANe!Plaw8e#{QXo=y0|UWP*hw*
zOQ}nFxp2Ye&A@(gAMSx|`OlgGidC&JDmF=&m_~j#*wyTOV?t?=r@YnAALk$?oL?Vb
z3!h&R&j1kj0YtMjX7`s+#-LG}%xYw6!pPYkHcLuEFK?Q(*lNLzkG;;jbel-dxxJi^
z5+`!&Y$RSuw32UxBnBf%K$f71uW26dCKh9Kst~Oe50NZb3qUlpA-K`}beETl?&JQU
zDb=iiT+D#A`fZ|2dY_Ah10-01hQr#=8Vw!^p94UxizIT-8!kZ2fQIvx2Ah!1)h`Pt
zmIa?U9BsDTL~+X4OvQ-B9uT)Ax0zdkQ;<>_;b{gVXj3L&Q{A=6av1dWgU?zZ-l{oqzDRjVKj}ZA&$6#HYZpKRg%q9*!a@5fC0~!=
zy9%fi(D1umq~53zx;*~{W$auz-6^_9sp=`y>(MZYRhd%D$$gmbe
zm?v2alQV}$Zxs`{mV>dO&1(}ICbG$;{-)l2-BK59B_*lnDr=k*E72%43r3G9)dK}7
z%bUV2U0Kq*->z8##pA68;Sv=JpDl!M-(PHT0~{h9bJ(I1<^@&9?-jMAnaGd$n~15Z
z4EH=5B?Aws#>JwLC0ML&%K{7w2JK@NNRII>3qk~>ZEi>Tgr^g}{i=a-mumRy&Yk3(
z*aaF|e(4u?7nZ$Q42h0t%~zoNmmSy`s+Wb}P2!c%gP_)@Et6dnXkwoC9cOVU-_YRF
zhs|=}ua|-32s!59orhUU^Y$br%Cy$Hzm#a;q|?z
zxbgg86{D)yuPbgznc^oAKz2B0p*rRIOio+9c41+gGje%I)tT`(K_X3*wGWu?JtRpB
zd|WO`=M{C?o=37+lr3Z1WXt=;NcAE%_B-+%l%mV@tlO$Bex{L1d%+<)uwE|PV3+Zq
zM;Voqtk_zg1&O)lGp1R;$3PO4I4^$3W*r$$|JMSxWfr5FV+DOetEE=<%hCkBTSuA_
z51T+RWwX4IUbSRRLrOekCK8xsNCULa6Y!2{#-qQNMZO)*C8(2wCk$T=iaRfCEQ9HS
zHg7t#L2nZvCYLGvg+2Xk5mRi*El%Yuyu#BoOV^UlO^aGsI+a<7!pQTeR=Mv{uiyRf
zVs)%gfNYg|yeix$Hjh!Oam8cn#ID>HM$ytt;1WOlcNWrw>(I~}+Xev5;3r$;8Tx57
z_)(f<9r(op4R5
zlu|O-fc;-1kRfCg=|6$L=JQNhvb|DKPifIS7A(|`yq)Wq&!Mc>#I#X;{YF+(H+lOa
zGGPuOsB?6`vMtgEWi$^_#x@5yQcQ!7V9R=evVr#0?&-=6qgO4J?fBI6%<*fxrC$}0
zIA6OsmTi8UbIV^AI5Z`EteXbW7>#~tgV-hC$eZ@wKlMU&SS-BR
zH^JtNGrS$OKMZ|`F6$*$&|_VB%xDvhmLrlMrNn<$(
zASZ0x3+DUD@?`Acp|M_mB{yN~CSq^?EMt*uxgm3KQ5!AW
z)|vhYdB6vRHC_c1dPM`@h#Up3=?!q@6kwYnEDWiYA`?{mEO=}*FS(E6{bU+P$>}et
z-NxJxUex^%OT2k)hDqH`-JG*?Q0X0ds_9SRbM~>iM`RN>Awa4>c%S5D`{Wb;J{o5sY!@!U_YkV!S{ZEI=
z202a`p^Wx2wyH7Bknek%+(_ghmv7oHXR5;a(zKOigmM*`AIA+c!X#{M!>lE`lt9OzlWq
znmm8Ch{Ibm^Bevf23OLj0Pz>L%-y8Wp3pqt`F(}$YcAA7K|;BE`S(ijV~#h_3cR=M
zn=zK841`q@p39GzojsZO7ew8pSW3j2OSW%noL|7jWVos_KACV_p5?ibbYqU}rMWTa
zhO6zdU)+mkhA*7B@#Nj#I#5lElBUm;?C%zFVIGnGCKd>j*Q*v_M`cF&qV+*H`b`Z`
z7}8V^s<7h9&J5ZaQXCo*R}Wq7A>L?sofw(9_FbMV6sTPRr?pBg21kB6MAQx0b$8${
zyAW!r7qu1p^s5&UV85vd!boqF-(wZt6*@;Hx{-fGwBTMF6U~gY7>03%4o|(19GI$u
z62EOYvmU7t*a^~ou0gI{f!@w-aK-@9J4mB5!4dK?0*vWnnZ-on-fP<0F(C~PZnR*k
z&hG~Vt>c!^EoIV;XP{Q$44nEGQl(9Fc>-zh^0@kSKiXT>1+p6DdpKZybQ9&Lj&3b}
zF9|TmPZYnmT2I78fcm7@KmFp%N>rhYFzLRS32+8$lt0*)thp^DuU6O(M8NCiaA-m1OEj}1RGz?9v9}4@l}f*>2-M|49KyFy
zXgV7${?VW8yvmJoJ3>!lJb1rCKDH^rl-bzNNWp6OUi36#-tjlHRd9sqvLud0V?sk7
z*8bEV;Zux)-DhhjpKZ&Awp1X7!Mygxjq6L4q~$qqw|XT!E1{#;=lb<{Zo$=ibqT+A
z0>kB?8~NO&VRLJBQ%?M979cY)^vYbE)2H!-;Kk4Huov#
z%AybhJbxZ73YEu@r%9vIgfP6*0Zwv3jj_^M)Ssnv1;@V
zUbnZ~A{d3W!N2IL#@AUKH^$e-&X$~%Hf{3c`Y-WsS`(1GgD7_TVeZ2Cv7Z-N`flm5#w#E0mNYj#)lp`aSbXp%0
zP79w7f@s(rWPtS14G{0)wD+7YS!s~%o-t;#2!$+3w6*s)kMk{Fno-+7b3eiQaN*Zs
zGo##RNYJ6=E23-l+}4hsKRuedY7BficFfx~lO@FfWTc;@-W$?nJ&PRQ*MM4G
zmVrU-)3)wx%Vl)v+RmmN&gb#R+ffa|`qWn5rtIu$W`E0In#51l%(g7z!U!6i?TPZr(o<97A?XnwQbHqbcDn7>JAM`c!L*Us83i<=B}YU72mo_BjhRhlkc*NOqBRCrTd%f?|Akj0>D
ze%$6mvckWG&Su00EslkhcH4Yj$sc8EdJOW#`!88SfJY^PeI�Ux%>Aqg}5&G2_;U
zWrGR>Yu2xjS~k<<1)eOK^rTOT6X+}ZzeVH*TH69uW)w8%tx^i0E$SxDZVa8+KS?MC
z5^jFU)TB#ZCh8BzmyKu;u4l2m-w^^{zpwKVFS2%}2Ic**2745Iz_jfQ#sUYHtew{G
zIolFIfv=N;n>(|R57}+>ve>-n=O`=_bcCWF@V8CFNft!w%}crVr`@*|c^~deyC;MS
za7unnyypFpxaWxZr|nDA^8Ppi&XR=+#)i)m*-gnb0Un>%%VV>1N1ikSM%y2H6a%;4
z^t>N%YCca4L%dkj&TAR|WozGe@aoRox!*zu0f=KQ0v8%dxG)V)rwx0?D+ij|^3zmI
z+*#8&^Ld;z+Iw<5_b&}sdSBN90nRD>6vKdTPL!0sWDC8zl#PSIWEL^5a8rhmCUX81
z^57(Jk=L$UwT;sIkl);Jl8V%d+?hq0a+}Y~_BLFniJhM$gdJjN)nDrmW*av8&-52?
zl(YTb`Nh-FuOlE~tlC>4!Ku;rSYu*+rD+<3K7Kd7vG*j>t)l*FTEhxf*_CyIYD?aF
zIqK}~E=(28K;eo)ue0yuWVn}ze9IuZHYJYV?c;gF4Wf|FQF#3H5;fu)o>pD*{TVCjS
z`sm$D*Pe-nTs@`hkwX)+>gtKFY_Q`Ehs1pdJ-iS
zE4-+KpLDE*JSibUv3fB6>)qS^uQIsv-C|jIeNCB%RMnD(N+USXpfdOL
ze(emI+EcY1L%&4UnI^2_vl(9eknWlD2vZVoI~AnhxuZnA($i4hsOeg?xP^!;|5(#W
zgUX@JKws0etnKhXYp3mQ(@r<>WRS#K;i}bUgns(Dq=epodZ7GuYsnj1WXZ>UB{aG5PWukp?
zqi-<`8LWn=)CAwzX0_>aZ?*kbkMOqcpxY*muA{rFwHVfy3iOgY%7iLWz*i1i$Zu&m
z$Y@gGx`tM(#n$zJudwA-4sYMzqDs%tQFPFuQ$Bj6xzL&?qPp?Q%l{H+#eVOWtBO-G
z4hkN+tZd_v9lGzxGb=B(;}Y9S<5<9=&}Fla6*B6Sry73Fnv5Qn6O35gyBkp%SpHB|
z7z#38eG0h?1~QBuwvw}n2Uoi1ufp-7Oz*(ykm=3j@euF8WZNjWqo*-)4&SW`*N8p{
znZ5-`6NXXT!oksrh@F$^|KzxcSXej{{s>V6UFjRPxZEhgR}W~NkuFR|_3C^e
z0U(BJ;aQ{g`y$p;hKy{@u~4;DmE>Toc0DRvi835ZS*u8oBo@xjbn?GZ$Ok@yz;5SEnhDML0V_0_Z>^yo!~`w$Ztr$$P~>lDfR
z<-6zvfHMk$qeb(~LQpqxt-hE@UpQybY6h4@UK!1qrMETPr~^1y`C_r9rTH@&oFt6X
zrf|HTpxBR2ruhX&>_8Ak>P|F>taD;X^GcUZAdiXRnv`B)@b>EU!<%Lek8Lf{
z%1ERGws~SKC+CSX|HXw(lMtS55G_iG)0}tg$k@%v#id&u7TFl8=kOhle*hu*p8l%_K;v3Vj+4W$z3GGG$<3KLZbm=Uarc
z0&^1`TnB9b!q}3VeJXM~La(7a5ICzE+Oz|{6CzI-!ubY;
zBPAtf0S$PyMi3r5Lm>ivjvc*kUII8dcz@mSQJI`n1xZo)Ho>>w#t!CkfC9W4qtBpv
z&{MW99t{Eq%3FMGT||L*2ZuT!-?%TJK6awyVYaYW5z}nsUNZ-e9mm41p8wz<*WRmXB;P)rWnPgg|gSnhj4?x{EKxR>m-#~1(R0P
z98tpyA=vyruC=Wn4~FdCnh#=`FIUlO4Iw+qdW*P&uey&fX9kuE)_#ki)^lZI4d7?l
zPvUnWFXeM5!!<+|0HJFq7$fOZ5n~=Pa)zo5^$da)fsM3Lp^`4_
zWCnJ1j>0Zpj@7|GWBMY7P?ZUrFpay=n$$MG4~$K3AGJCq
zDG-a&8d?b)EPpm{P&gwjna^^6SFIN?F`Wt>Kkk?MRg8yC3=F}YNl`_@U|72XG{Ui?I-FiJDD9S&s%!2QoI;^i`r1yK>9Lsd?<#{VUPc3S>;Kt#7pYT34dfV6n=;QXvXCQZ4hy
zJzE$1ANmTzEmbL{3l2eA3gg~3-E2nYE2AfB)$w!u>UadQWYd?Ru^}o@S)O~>R{o!f
z6&I#k!JI&mu?(&7pe2_Sm$0SiIE3I~Lq$ZKuni!(k+_~-kj*$}8VSdqLpS^5_l{3X
z=GHCD7v>7oK?VnoBp(;K_$^y)Qe=easoI||zQP+VYw=>}`t4@jNVd%f-8VzBChc9x
z_!s}=A*w^|;W%zrqJ+rqU!jfnvPcYpK~b>bi?Xy83_1(rAc?jij~|;I_~|Jfin4ia
zv!wt^ob-)H0#N6`9c&8PdR*#^?NR4Yv~=-D~3>79#5I6~K#1x%`U?X==f`YiWmS>j?P;kex7e?AmWHaas^tC1%LHjY)xbv
zJ{&!?6TJM;vS_+i$J$iVd!r2AE@w!XwKGUKVf~t;92R9;%2~-~fR_xTyHh&f?!YvC
zvI8tLh~Bi)z(5vcpdaIIwYIi7?K_N?z81z<^9qXTSm#C^J+0QH(ZqOrzLsipY8{ZA
z1P{(ZiyPjWERVVhGtHLf3fX$5_QhB-=XbOq>X@%CFjz`K0!|g4TQN;J<=1ft95TO*R&NLenKB%A%rx{=#Ll3ZHITymSs=9FA!tQqMIxknl{93j4=M1>3%b%6l|*_%WNc@+SSoSm1*H@sm-6v3?Z+
z^WB%Vzdmg|I$0}Fc)ktl2??kCQYwj{-%52nD{DXa5Dh@OLUrwl6mK{H?D!HM?t+4&
z;i#RV9hn>g2hav^B-{?FYf17U?-2GsGK0IEGDc!daA>RnQXgxK-7zyHpbub8kb355
zq$$<}gEEBwe?)R;k+nC;H^dK)LSixLd}9Pn
zb+JH)D-UJG*{v6_GVaM+X{ERatD?b5zPDiD`Rq!+q5(fuB~2@LPu?HNX5h{;8Kx>K
z4T+LWRPThTY!ANDoyl!kYn*!cZ7hleIkfpesrH&n#GP-ja15<|g+jy;Ma7Jx*;<|ycRtAISB?Fj>&&IlP
zp{Kczxhv^z)xw+73zI?`6AkZaqLJ>pvkcO7JRz=7cT;GCze_pR@R%)R>-NgT_#9FK
zVck64BiSa61u>v+&Km~H-fi0}(E(A8Pe+27&~xS0For!+(%0RNPpM0`X(MLex{hQn
zzO78|Yj?{2WMk0oBdOcg+>X9#NCx$Fs2*(-eFo_{%8asou;%Tk3dX?>R&$?yB
z5@lUf&u`Dh$E7%i>o1w|yiT#-;y=?lPO0s4qKsdf8X~b(3s#)CAM4|MtVw%QFm8TvvtRi1@!KeR_Xv
zv{uAWsl+Z+UB@VzjZnjfnvI;AvwxYBid6H!Jqxrr!k%G@^jVZ93Gttx$_u$?cju;R
zKS;?TuQyu4gK@=y&R?4+g}rz_H#v&NPmY&w?EDM2d_apAoqwx>*^M^wOW<+N8YXiT
z>2=<4yL#26<~gQU@1@;)-siezc8aRtW%%*E?5W5LkB>H*YQ(S)gz0`GiR^FW`#MqIs%WOZzny!-BZlaEZg+c!6fd8C=r8TXL^)Qi-ciT6IZAdE>$$(c3wpCFj!QhXz94#mNSR=A4W5-@>YW
zGNX>o3_5wMoaPVaUb9z;ykXNgw7uw0Xp~Y3e_Al=9FL#M+@SuPSIFEb`B0%Dw_4Tk
zF$Vg9Ju_%7xB0CE-_t!b`dy)pe2|6WbxxCO4i_*|C2`LC!$7?W0I=*91YdT8hsLnW
zroD@^QZcv@81ZUx(oW}i*ePcgjF3$ZT1|+wpEb7
msPdh&?)t0pel_0p5RYpi;n%!6(i&@jJ%gZ=}hj1oWq

delta 114732
zcmV)0K+eCs)EuqB9I#CSe+@}u^VDRziUEWK*Q$2ToC3sVuzyh~y`;>4$i_*iT*;8B
z8aFfifi~o@b%MlJb%og8gWUUZ4xcicw}FXTA{rT3VF2x4@7RRDOag62FN=2AACB|c
zGXPC*6Nyft}>X=V9DS9a5RjFT^g3e=4z6sWK~K&&?##
zWa-nLDLGuQMl=xnu=S%v1YrJ4LnVll0oxbfd+6828?RyB4aa#Q#bSgWH2Aho4GqetW(mD`>>LNn#ip#rQ_z(y6R;XOTBk)U*|&6PY?4j}TAYZE|PQ?`DB<;l_75
z9^}d}3I948txor0e;*=+Mm|qr=!TcjiZg0*20U>uo7bHmNj2V(lgU<#lbK?m^tR;*2%@Pb&FWW(HmE!_YB_Ajrv8sKQ(;1K@;Xp7tc
zJ|Fv*f+~`UZdpPRY3U>3u90%a!2=Vng$_l$Lm?}ci0_z7f5n%{d-#XT0+S#GP11E>
z_s&g-FgdFb|4(5aoAFJyjCI`8f0C4_$${v2GWLDE%_WVr01Ma@7j~CCd(XNHD2R*n
zt$%EdIY9Io{bBUf%h*d>DK01fpuRK$;w1LrWygqe^!d&eP$x7`omXo$DC2%zRk5z$
zREyGG2GULCe}Ny&h1j`HOl?$cV^}?MOt>l~dcGsMT`4AWWa2AEfPfLqGoH6IS?-no
z#ma>1yJm{_<CeLd(i6AIU>y2F5k2u!nCD6Eh=L~7b#4VB
zmD50?f8+K|KvJ({eV_n~y!|xA4ss?Gordfhn*|mfU}NZ6s!7pRb}3vw;*Oq7()*Jt
z=2ny3DXn8(pZc1E`N5j%+s5f~$sD_Qjy5^`BxxMq6EBb97
z<)Sf2iN*lzxW*A3E`ort4en1)DQ{98AB|~}e>HjCg|jd9Iav^|TrZc_wnRKCe(1Vb
z|1Y2tYVzigBx>dpDXzcAK2Rda9Cwv){#Z*m;?$#w3&M}~gBIizL3&b%p
zdt91rv2;G}<2wDeIjdYQk~3b@wnrXuj}$4B0D}Z=3Cly5^8ivVWbb8nHsxaY;#S}c
zf2|>pmKQ!#v=VvyOY4_8M&>c|!<{f&~wFS80@)y&$63CdjN*bzWSu(hKUvo4d+k^17rV
zjhL_o<|tt_N5LQc-3s-pyu>uhV#wWOYWeNuQ>I-+yd(*KQ$dPSXpB`Gda~b3A3XiG
zg!R>LOH>a_ll~QbqLBuNIgeUskF4Hc(vUk35AuC~N3{|S<>^Rec=K`SB_Gh$Q1`;p
zu5*=1p66B>sJ4+oUF7OQOKh)f^X=oie*<+vEkKu%Lje=Bc?f}i0yZ?05t1r@8ryE<
zw)Nd#p~o@OVZ=L`#R5T!Zi~WsDAqs=Bq<=z$hOk)$dP8^Y`?ze@S>q)N;8{v`Xq`v
zB+vaumBm9*7Jt1e!}Gh_tFJ#$U07ljQ^oClK{+Q{nL>vE=;jubg|8)EJuRkzP9C9Kd*rlpU$5y
zOTg3`$`MQxiAPdWNUBWZ+^}p#fbo>C48^4s=$;2q$m{$Qh>l(5g8Ze
z#HwI-R-n`Zhri0;;6{;}yUg*x4Xxfn69FQph!ciTAybCD^H3u@Tr(y>P^dk?0i&Zl
znOW7|_}aTe)$Ft82V(($+DVB8eB8nxORBPTn^L}J`?@3?kUP(Nt7OXi_ZcIgC6QFY
zgivzc2EeosWv~Guj0gKqmNaK^NSWamQTFA&g*zo@Y5^-wBx409WhwtZc4x*Yt(jQO
z+MY-eyAKj#Bipkh9Vj!R25kkPbprKcDa*`}oER-40ZUH*waN*9063>|unrt=qH=%+
zok2rJG-m7z{3N5!z6=elynSqXZ(Z@(9y=TeIEagVwez#}FURW8w0q*c6lf!yRTi|M
z!VnD(NlHT$i2L>lcKzqIk=|T3JR&mUSp)z=3sC0pFgx_PKg+Sca4L*L1b@<6T{Kdko`y$0Grc+l?hJ{+T~qaSPl6<p>!2UNxC1=v%^?Chk(0jk-1s6ncrJWm1U-hZn0L9~(8865|_
zIM}cPZv`oTiNotd)Affx!BQ5ug9k7b)EEX%)bR92!=TweAH%n8b*OSFTR4zu3#pM9
zANcQ#)q@jJoJX_&ZAMArt_D0Z1e4LOUZGPoD`5Ap*HW*7rp%+sHOBOdu$0zJeWV%%
zFC+ufOJRfGrpvT&BVNMER&<1>@i1<$~^8P|nyiO4}-Q2i}f05vF~
z@$y*$ghc#7t^$k9h
zf9Uvl=O=G6gfl}d9Te0X|C&lDHnA*Xl9;HLkX3*I-XEM<4QY~hWvD-(R>5B;QMWyR
zKF5e$ciq6!kx6NRX*u>G?4q0nh;f~OE!IOQ$$)r{Y_MJMyr{SHD~h`3%*dRt1OnudOaNPgg(uR2EzKhLMni
znZ7v4rvhT2Vd!pl#mB3E_^#rZEnzD`35%uZw^=ulpmYg~mt32OU6`9wFNu|Qp1Yt5
zN~DdM8SbJCp`52!-yzsHGszj9iJBZzd$+49vxUQer|XbJoxExh|u6i!P&fT}H9oW#pmDs7L+0@n*iu
zn1Y?_S?ag#S&FymSxWnMP0KGtDfTQ+C@nTE`Nd63eo@nspA9MUvZ&4HnwIKqnwCRk
zn+Ym*5>uGeq%<@(VfK)fLHQ|M`N?&Xj_TF%I+mbYA30zy#21~m_e^Mi@5MqU1-pYi
zT&pi^*@tvczv^)5)wb?87=_Zg(5!@d>wIjA)e?Fd%G7df8rE2%OEF%v6q+bxHzlMC
zqXqfUBsskx5X61Kp_Y&}jrLuNOjLv2>It<=kqN-1{S1UGlvaF#&{AMx*gSZepdQmCeBxXr#Q`(3>w3nV&m(4
zHg`9dP?m;5s4z~QnQSP8Jj*D5eWw?rGqyUDAo$J`Z7@hciH=mW^B39^o|QRvjb!*L
z+e%b!@~?sBmt3j;4h#gPze|a5e_(Iv3SVx&pz?*Bj)E#+KNJ1XLg`Btgx|%|H}0h^
zM3A8zUrUu`8)`d$HJ4Ve=|NDCf_hp|4>N!53ArenTvbv
z)9&|m=*8l@ySj^gZaK*;sIWwpHofU&$z7dxSqtV%n8~Dn6M)BK>nJA|v-(@@
zZgwF6H37)ozQH@7gz@QD*Jj=)IT8k|*25nAu($`3aT4Zyf>iZ~X5+t&QmjA|o07ye
zQJAbFNn^Z?artjwhhDg`r@pwcu3_+7_4K@Ru@Zpqd}Dg&FRES6>gL03D2$$(eY3u+
z4jVk;F|IpdeRp?>@NAF1KMKR0XIuO1U*jKSuaZ8>BvCeqra#oreY`*F+T+8ce|f@l
zk2_>p19JNye*a6B6S*8=Ctqoqq`KK>)pfQaOTW~G%KTma-^~%ONxx5;1nfgO;D7TW&9XrvsF_=PEtW_So{VCwKoGzr3wp&Cqt6v%e^qCkV#df288K#3
z3^hD7W6TeP?@%=OpgiP6V-%L9h)EuDVuEs_y|Wd2j1f2w{i7(YEG8(Aty#$N*q6o2
zfGs&gDgfs&DFlous{lM>Ljib>u?oP89v*lPa3t&u`bF$ewi#KB6krF8n1m%E7IuLl
z4D`j;9YNm&GZAx@e_%s`zEH_o*e@~cmSH)V%HzKT69k)KLX5B%RC1B<-eH9DurnLT
z%MoL?6{a%5uwD@R8DIk@5Z-ch@+ROR$s34EST<*jkt+ft
zyurF~9&GKzdlp89#Yn6uCL@CR#b#tO>;v}`Cio!oNGuOqe-OCjlND!+4?=`w4=0ek
zu_}lRvaEu|Iv@ld#G^1Mj1U*x3{E&B#Pd-eFbn_i04GHuE*1{a`%26Zl94xf>p5f2Ll(#OJF1^xs9J>JOtoo2q`(&bnsSU8*Fws{Ytqwpa6UbBS15y!fe^
zOh>QV8+B?V4i|K>XBDQ;F;Cfr{AwxrcD-MI{Z00-qwvgl4e!ma&d<*dA>9HQLevCS
z6hPULc5q7p)LE}Bf};RR(P_82zJiU;+xf}GXxu3Ge|gG>QP(x|nX;dM{r;wVchZfz
zM$zI$RUgBwN?dU)8H<)aSa56gEVNiXi;tJjUXsb`SyJ8dnc(BN!Xx0nUecf
zGQn)Sl8HdIpJif8^@n*oK5062ihaC2R`sXmrc-OX>?H>j++v3vNDrHe`LPLI62wN$
zND@U@e*ShtvGigh?2NYyrlA}o$A+Qr8$n1{Na
zJ~bbSj)(HLTq>;cHu{7ZcXIeZ`X;|k-{gwuo81%-THz?Lvlj43qC;nEKUh){OAe|#@LU;ov07vI+Pc-l3WO*6mI)5}g@TwdvBa#jC3ZLTNnxSq7vv-5T|`EJrq
zUy#u_`)J0gCZ_1Vnl~3O@Izn8axX3>e|&=--4Z}9*h*u63w!rbYd_c;gmf+g`;@gw
zsSxZ@XiOp4W}F=OS&n8KX=a^Te}?j#cK9uql!FK^H=qw;I0f`9At-Mxl7B!~fTMAg
z%T!@{x*x?a!SUXePGL
zamqV*vg?(oe#tp_;_H>Gei{36S2LKKJ9h{~eppo^3iojCzN-No%Tp=@f4h_**ggBr
zTi8tARnHcGs;iTy$e*$1`>aqFKYb@w~L10lB&mi@!S
z0msQ7YRIluSo{HMc^5p?k#4sn*s_v9%j?>n!QDNQs=H4!un+8WUYsY(BpY-N*?$4-
zQG0lCqnk#)FGyj)q9~H{{>*TeuMR7|
z`pXwQKL5P^;;ZjOxvE)hrCn{WR)TS%;2U3wmCzL{gZ
zY`xK{F1FKo%?;A_ZF3EzWLe<-PQWH?GM0sWM3&(+G=IBVX`9b$LnUmMPGdf
zGc&1creT7wwtHCgEv;DSrDT;-X~AaKx&dWYstA9>=2xz??`fuqUryZ`IJ~2mL-R3k
zexg@5yvndq3TDe{wNZizYl+t1<2toqrkt4ll^ae=l$b41tSWU{^9G1=TYShIB}&FC
zv69Tnnil^%!`Ozi3e>TgE&G*AnK@6h+UOZ?0?kwvujA5sWm#RS3YSWdQr9aXnKp_x
zoT7gudQMxhnk%GTSCRHFnbUwpUV-l5R4L-RCa$w{svt?773AKP+GPkTX6z&!&)HirZ8bhDFcQ|i
z%NSe+d-LDxf{Qa2E_k02O0%j0(Sak-02zO5mmyU%tAXQ(Rj{SLFHd0HbDs)oIOcm|(&A~M>-0wo{
z<5(A?yP4ipL_e550pj_C8xbeU^cC=PbAfr)4D)b*N6)=S2s*`m-8o{>t{X-=e*%AB
zkwjSFg#gYgh{lbCa1#r!Ul_x%O4>`qgEDRXc_3o{IEXCyE}1Y@CJdJzr1=!{?J1W4
z06moc{SN;KbIs>s?*CU?YN4%sEKaQDU07(QAznAYU9Do^{Op%*s^%FToT~SS-h)R$
z1`KH(rdq!n&oI@8q5vJQi@xaSadm%g+l9|Ix@mUp4tJqa4pTmltJb}v<-CRGG#5*WLCS|b~+b@9)2r4nwq-qW`68qt5`JkVS3
z-Tr=w8|UHD!UOP5BhU?&VWb`ufrEDl@V1DUE-LMP03;8F_J4ZJOY0q~XAW!4n5v+t
z(I~V3u=1V*CMgFF%ny9C&8
z{KvHd6mGlQ=5$zZOleX5rZj(`bu2M~63qsm#`%3Wtvva2@7gpT@ua&o?CJWOfoAU-
zFh>}nQ^NO~Nvi<5I&v|0(Z_KL39GtrB&}D`77R~f=7N7NpxG@@9GheFU%0Yl3B~}|S%O;(?tZFmCg7OK`J5Ccd+6Jb
zYXi~NLbyiYKGF>@s!bQS;nnvLdq;QLrAjB__~?_+FWSsz8TbZWbpPR;jtRhdpNQVN
z9$<0weRK@00Gg{=<%T|kw^vZjmnoH`peQjYW+sPRk?cK7*9(7S+9ZF5K&VWr$0pAO
zGuulMpI)}*O{M9f2f^&xW@MSO>`1mFsp&JfhcX_e3bhvWn>
z`0=Ojf8NqAWC1V+)nWOfZ(!mj8J?Junvg0D=DLt53}l3+br9z1aI7b~^Wpk&?oUE!
z&{JKeOTLdfm2-d5(-s35YQ=KPuxoCbaTA~>ueOP#w
zkChHDejdT6UB4gbHNsWE8wmjxoafX*0eVkY0{n(P_TCNgq$X{pN9+0n)n1vbDeP#-Bgp`zR0_02f?*s9t$hh`j7+scixkriil-SOC+f-l3{dGN6KCSF
zY7o&QVb(n*lc*{tWVFQK5tU<_HBmjL5M^D`=s|gU`Wlq?_%1i2C2bAYM7yL_<*vYK
zLuo@VZ?OSx#PS6?6-984&QqnQtTp3+DT_TjN>3E?DM2YrP+%N%);c(Ax?n}?!h+D0
z4WoY)htnnR@{T7!#?TrKcRU0JbO}=zSLo{DXhR-O(@>;f;B_<~C$Dg!C^it7AuSh
z@nL}5a0$cWFuWh#*f*bYJ&a@=W0{2nw;+G;uUsGEfhlyhvb4)RuaTg5s=X&>{!@wz
za;HpN&Y=H1w+%1Q=!yx0ZMWNlTvb+NV&rtH&C0pVrtCIcDOo$`5l?leM{P`pScjh%lv
z%?m}I0j-zyav#EPDd~P~WePtFHwQ|AFz0AunE_!wWWcze0^CZ%@gQqBq-)Q|xcouX
zGKfk?KPxGeEleWp%{hh`FgyeSKK=7DYs5hWw57~Xg`wt5@hJ?cdX6f=i$AHP47cV`
zZes=mBb$oH{1e&(eLt(r2f~RyB)Wg9Ocft|yNPY3NXHG&D?`2Cpo(2U8F9mmDB);OpV%6KXm?g?^f;sipip+L==f+NnI+
zna3^fi3yVfO7L_swXMI-Hiu_FY^ux=;TMt5K0*B96RMX_2vGm()vKI|Br7w$?u9=t
zdfgXyeT_(szpz$;B*|nKYu;|Z_#afPleCi&PZI(+Gne2p0Tcl>m*L0-DSzcz-EZ4A
z5P$DqA#_;gU^OR-qCQ&mVcpiE13C;_9yWAC!C15{M4}8yZsY#-JMvDhtSE^;(r&Gi@B4m~eQ)9U-m6DGJzrcr`tBu{$24um^ryq_h=2FV}gI5l?&=
zy{=VZ$03Z*IqZOKjzd*n;hJ@mn*|P=_3C_TPpaSNjn2;hzW8}Z3n5q>C0-Ot79^-d
zp;c`h0}=X868l`$I?HD*PF_@wX||cZQB8xi*Cx}~IQ+|({);!S;eWxAFt4i;N2ONH
zy4IxzZIems_VP(J$G1zXP`+4_B|2WYadP`CO~p`fc#43`U($(SYh^M!h^%FjiFP~_
z#}i%>PdC0=*i8(q6(*C{b0p#>950ofyzfy)bO5T;0S=TnwPuGM6}TS>?UDv_tyfEvN5Zf-8GLD>#cp>P^ANLUK*xAf9y1AR9Wm;NlBIF7CoLY(BG8>^
z;fSy)<0A0DQb|mw@)6ZT5R`WsJsn{aI7`!lM4ly%i#Lr~9Djnvz$B@R9*Td=&mxKZ
z!ID8p{N>&JgZuc$-GWq}abMs8kj^sM?|}Ukkn$`HZW`#p_($SQMk8@2e8wZ`Ny*}r
z<5Dpec=1S(-AjuHv|ixJ5seZiAW2e|M0DsgT(r|CraTRqhyuttELv_Q)g;lCf^gER
z8VcS*V+{$h?tk1!^|7Q*QWiwQn+AfVoNV&vSW^37VEHMjp=>9v{sm%$YXDM9v423{
z=kaGF96}ZbTMj`m!orRA79u!Zi1-#E#NMgLPZoloE`+?Vg4`y=mP5c#w}k7#H&bE+
zI)$lVVMx9E_uR~ERwR6))yx8bCC+lQTDLST)xxa_CVzQ@B!$!9N1S3i1-vCvxrJEv
zh0!|0kpncaF)mkwA2xe_LvK_zrLJYJi+j~4i(
zcx*E{?J_<^LE%Y#-;Du^1S!e|GldUM3ssFQHyyU=Rs!-^*Ji*zQ?0s#X%BjIBCs(?
z7;-O+Qhz9{a
zFHm?(ux7Wui+#YY+Zcc)iq0z#cg|Fiw}hqLfhOAm#CU+BLcr&ekZoIoT
zH|Q@Jfu(Z=G3cORJb6=f5N%cgvdU+wbrEE-Wq&=ho0fU2XV#nq02{E(C#j*DD|+!m
zMTc07Uf5M6K^Id)kpZYh9j5(-@IF*~7`K)Yb_v@#t{VFy1cMk{(W)x~s&=L_Pz-Mt
z5G^jIE!iSfyUW=)wbEK@&p%JBl~e3jj!&&|M_D!Q#({@cWm%bGGYVcqct!+X)e7vV
zseh+^SZaeK1t!>h#ES+@o%b6Tj*kExLC<2+$DbPRW4n~!UfHV|51}2Wk?q7#V4x?5
z&=GmUWIzr4tMjQ0z{%yRs#_czJ=0C2>J3g;YIbGM#bdh*AoKn1FrSX!q0sTNcRU~m
zH^GDH#1W?O!*`i>6kzJTMH$!SYICqew0~JOM<2U7{7h0ieh9T8_*O;L7WpMxZS3gF
zn_yM6N85m0@9w%s9^YKK%lEN*kC@+vspy3%;8GrLo3(8hjYXfwCf$_yK|x?in(kzs
zHjV2wcI;}r(@>Q(f?Ax}uL2bCm%46n0_6gXKs?^-E>qhacHX7ht{b{wW}%h7`+spt
z?NFTcCTxc@5J&)4&s#az5mbkF+<&VId)2DP
z{YZI!zO5|g9lm%%U^lzj{n!~
zuxXwmA-PU=!Xpaz=^a#*yC6HQEFTa`g*LZWfG!9=UxRyqN>l4DXnW=NK-s28r0|5_
z0qu_do)LecBKOCAz016#kC)B=u1`H|!J92B_ode9QH|Zk`*%Y70W-OOX*>jWX8BA%
zOj@3JnTN}tS?C*0bD&HA-wMG6OZ^B8n6e~E`hU@ExBPVR=pRv^>yrUm6c{%*ATS_r
zVrmLJJPI#NWo~D5Xdp5%Gc%K+*eZWpkK4EvexF|BzF7?;7#GPn(OUM=g*eNioog=%JZH-ET_h=_lJ=O6tT
zjC%h1!&PuQmrRDg!3l0n*I8{(5EGJ-2&d7jL@d@dSo4oaM$WExGpkus{9BzOcxc`W
zC&{rkpa*rUMrNcmKamqK=m7!7X|m_ED45fPWSpV9R<$08Kk#u%QkKChJGyXo7n}Ui
z8y+XH7R;fY3WJ^r6$B$Zjyr#u)d-ZA*)&j65XzVrsH)vY>AY1FxJAq2v1prD1%oxd
zxg7@jopH*w85+!GOhlRnbE9B_I&I7LRcki8hizH6#)4Di(6O9%rzWI4lz1|!+ZBbXRsi!Y+v^J`}_rJbDQf&)sAZO
zZE<(k`F2hzBUut()Wv^3L;Wi6KQJ0lN-`;#0c}n*k_sNoSwhk@!$ogqnSjFy?0<8f
z$PiSHwqyEKO_Kt7PHuHc_cDkjRIJQbCSeqAVo@O#)F#;2>;8BC85%=
zy>E)@5yWu;E9B+ADH_v~;lpfhhQJ7gueZ5|P=yTaCbVmXW1W8@*y{!4*xbNhmWKa?
zgM@7>t!tCmoF^b4cas*mO>9tiq%@D=?3kj`rOFKkoQG}g)~XgW8rE8?g%?y;ixT>F
z4QSdB6e>t2gmJAw9MS#zEajm>X-`F2;uGY((w}FbuX29kJy7F}u&g_nq|uuR;29+{
z?Z(tAlw0qd^-F(Htq;};VF8Pbrya7~I{R>aYQV^Z7tW3=&}HDWe25wlkAOrZOFhsK
z6$xd5gs|t#0sW)Vj!JZ;`McSI6Egt11wX-qd++>bpA42XwswGx6}B*FX&1D#7qqsO
z+vqagIyZu!+(gJwkOE0s?y+N~jiWc%FmPJOPDFVW5uSgzD0gewXvE3B5&}?=9`U8g&j@4*_IZ%xrk->x7?)89^dWFbG^eEXbbeV`6uVeINYe*thP~=
z8H%|&r8k@ip>iguH--%jpKO-q?HX)X8vBe*_=NV*aH|#6F&j{kSGzaYPej9dBNnyK
zeG2C|6mEaQ)7sYn=ElVg$$`yYsub#kl`&EpGOh6Yo`^x4w54ekE
z%xxZay@`7h{LrCb-}AirHlLu->WXY?aK(M2j2+#<>x65V69>fu>c*FEkx*FyYhSc%
zgxnjH?;A)xzxOifaJO%H$nR=dr1e-`f}kEY~Otxm3pU#7dp5paK~
zwxg9^5*NOskNfomVNl(pwAbeGiwih7=(tTsF2(rrzFF^9t5Pj5l5i)RJ;6%U+kO8c
zHqshBb6KrX=Eb${r_(P_fIGyV!=l6R26lTYaZG3;k-Ygcl1x$97%r27E!bQswVbbN{babLTraV=QagX08#P!4
z&#j@l?0{N-!U3K!jtt-27{ar0{2NktQHXr@`wy8oq9)G(-?5su&V9u`0R@SPY&fxW
ztfXPYkX?@~0S^W=P&!IAp^agIlOAYHf>0?4qZtZa7t5upCe#_sj3vyZ!ta|<$-^7u2tk=eTYR3$B<#K|Lg1!-
zyNECbV!l?{^sgGNZY{elMzo77&tQ^Fd%PYY6?6&
z3NK7$ZfA68ATlyFGn3&#D1WV4ZIj!!5&nL^f*+~L3Bel)u$g40ovG_~oM{>>)6OLM
zKvOu0HAP-Y>dyAB?*a?(O+3j?@`)60q|T1C5k^mF`a^Xu*(ASfn+ifEC65-0Gj7s#TNWPeI%E5H)MWfn0`
zWFkY%j;tYSnZ
z35q?hET;$)up9%~WWL%|mmq?vXd;xasDf=s(=S)`jTM63Nr*^Qv=xk`8ozv8iH0VW
zErT)hnv#U2LE>RWLw`bGW&iybn$T4nwUR;R)2a@t_(Tc2rQ{nF0M93;5=yNAjabB_
zCW`SW$tYnkZFjQqI-1t3`4}bvgj1f(Cq3L+NvA|-*?dyjZMH)2xIY?v$&0>x+9XMw
z_hnP#*u%Ll?63PPCJAa$r{~&V0*gYh$Qf8(G+Z!ABwcvhnt#XEbf(^L9rxaevVJ`G
z-sHA9;ahFaZBt*dTmk@@5{rGgVUtW4$@s$EvK}|nLYqRuhs_s)D1djOG^V$e2M-Ff
zFYAMyM&l~q@C>jI9+?)0@_K*8a?UB?ki;{kH*)>To=?o>U~UG^&NL%dXut}ErL3vt
z;b`sd{l^BM7=Mc^c;Ol0$#RA3R>9r^MX324vH=F6@h%wU0jx#<)3|x?#%;DNK5kg-
zWErqEyjpm4!DtsZwhUUD0%(1Zz4u~7n{bg+)5r3;l#--F;R-hVX$2x%2Kz05cDI5O
zq0_63F0;>|^yC($vl*ptxzeG=%qls9L`t-MP?27&x~Nv%mPANUDC$mftyDlDH5*F>
z$WsbNZhr;m@pmg~p+-_Yctf#(ij+`Se!#nr;0SuB(s-45?d_|&2IrkS|IJ0|B}WxJIUh?R
zw|~B#foj|E-aF>xS~}*`fjKReiv*|)WNoCa@RH8$7+MCi-`R(=7e)`JHh{A7liA}}
zC`DK557ZC^*)p^C{f+vgA@RAK&+$8tOWG)%J6F-INyzKYmQ@A?-B~(ppFJOd@)R2P
z)Z?V7%{D*eW!?2#n?MkQYTd>`52y4}Pk-09d2bFMZKUr9y2k>zN8ZO`j^B#Bs)nM~
zg?cF=p$jW@EPGR63A^r-&>V<7o7-|aSAFSr4U{I@<{xwU$8G*GpMNa2Rr$KBE&-Et
zu6{_sOf30IW^i`$VSC;96auH85PtWbWBMuA0AYDa_oHY|r$9#7>iDd%tmG>^e19?R
z$by5&_vJKn<>Azn`}qu?&PQAy
z39;R-Z!Cp|G4mUX+}TuR;e(Oa2Y;NppP@v?KSH)+G4il^Q@YMM_h|X4E`NQGYa#KY
zdGizYs_x^zH$ZjG?SP205htZb)IXDF3^^}ve;U@r^)ZWxJbUt;AxjTFOk^g{vn%q(jVj
z-e15vN>2LcyT9)4{&zPXB7Z57Qbhs|3`xqO;&k`(uQb}jN1L(&D)+M!J;B)#dUo5n
zRngnKfB9>{S!*41B?=DJDy+HX(29R`bC-N|hc0kDb;EfFR)mgI>EI%zE2ncfTynMC
zrx{M*v`2j(><20*@v`rH-8ZB{jqJ{
z><6Z_{eVsUru{hPduM#gR?VU8`oJfxIY5J8n+HhMk@E*qLF(>8-RSyN%7Rllq0}Gs
zvG`V=w4v;JDm>Tr_zi{3^Q)(}a60Fd;lUZUl|{%T+-(`89@J3n!vp16FmP1ZcEt4_
z!!0sImQ7K
zuJV*0e62X<{w_>^&^Yof{!ZoOX2HvZg5)nGAXJ^NCevo{=&JYAWF)T_CWv{7UyUTU
z-4`Sb_O%Z<4Xwh;?b@G?e9pdeh#^zIVdGy9UCG2`6&rUr34ih66U#vdr@(bU70z&|
zW+L|S2-S>*4v@#dgOf|g4!=&X1>MWo8}ay}`Qov?6)c~8No4D{nuD&|+CD7Chbt^H
zbHfIJWs=ZTPj9hgHfWc6yv0&@7K4s*fFDRs;0MCPHIdVS&puqO!TaDNyia$!^;0^0
z7ESG1iK2JOY+RKzL7Bmo03&O7VsN%#OoU@5c+oAfD_5|a$`2SMiJ2i{AbvW`RQ^6Q
zTq=jIbJB$!OM7TqZ?Qigz4|wx_n&xho{C3;YNGA`Mv4;c|Ab=MOa@Bk{{i>&h5M5c
zPZN{A%@qSVH8q#P0s$(2OLOG55x(nJuqqd-vI7$@J~pL;Q)N5JDT!5OSJ}=MHA88*
zJ3}fY$GhuapKf%6B)E{Y>s)=&fFRHf^z-Ycvg9GjlJBlE{{8Fi)mPujBB?}`t30{A
zPcl)J@G47WBDEGpUM9IMgwi&-Jtse+!&UPj{F0bZEN=e6nfH
zlWkWIA8BfTs>iYYIn$;YUg1=gwiCSqN@dz#P6v9tyWV*&HXSdpCEvGAhl~7%b~+M=
znpXkm6Q87OE<@j4|8o0}=(UU$s-#=B9dOERIvpDt^q1-JGU3>yry3c4#6Rx9s?5^&
z$99jmKGO0%oq4-|Gvld&ekJ|X`+YxE@=!BN3w7=(<$mgqXD-$;kmFGem9kM;v
z!?jiEfwn$nSE?*U4w4VHZjB)Y`ouc)V*CdvGP;7HH?{nT_cM1=v%u0hf4+xkQ|+TJFszy?^9
zXi!?~^fKaqQM8GwQWqkx6cD*Rz?Km>MwBKiVRh~oZB+y;$PG$#1AHXgABij6Sf+e}
z$DzNg@7fbw)VUI^zNcTlA_)tg(HU_Q@&gh~F5^uD{Yv3gtGp~sDGI3)TNT3OB#LiQ
zkF=$75ki-AS2qDc2|dllC3xK4&&Wlyu#bHlhb-BDg`^KUFy^2`uRGoVjDXP8439%I
z(bUQ~)kd4Pw8CbFuHfT|1_Jd+fvo{^Y6hA^=@njj@Qxe;rq(hqoXp{^lldpR3X(az
z_g(~38Sh{|0bLYZNV-7MPefW?LA55O3+apAMEGZdTta9tW7YMgp6=j$bQ4jVlIFHn?Y=RzK&q9x=(
z+7@L#2gbL&M6Z@Wc!9*n3vSSnN^q7>NlEZxzw9p&Iz|{(`UNsVz<~bcLSZNKB7OaT
z`gM$!EQGW(u__YFk=T4KmfJ1IAn=e%Xls(4l%leh7S((7JPXhYt%GFeZF%J~fqqcB
zAYAW3TZsVVDvOP5gHslw0{`L#@7vQUa2)mQq^omMk5MaFtMwo%CX$7u-B1~nT
zjuJ*lh~BG3Y@JhJCcv_-W81cETNB&1Cbs`@l1yydwr$(CZ6{~;KIgvO*M9G=)m3YK
zC685meV(yV@uUBCTS_dQXqLSSC$2;G@vx-OV>$+`ZwxXkbkL>aB~KJ_Og{dc=mU6S
zEZ=hRKTAz66Y7`#H@-_x2+D6c9KKA;@W3PPj{u;uf;Pix;?!O8IeB=oZWN)VAw8_U@?d?MK
zn*u2r3+$1X?>R_^Gz#)w+TB!|mjfa99f6cXZNdz@&ES>BB5^%h;GMoej#0+w;gm{7
z#BxSxI1(=l;OXJB&lfgO#qz5=PN~EWjA)gXu?TSizhOof(ynMhD=gZh-BdyVeN2lC
z?E|Beycqum2=5N5*Fy(ZJ)U4Vu%efcCw_Y-+ExT
zVwl~~t_Wzg6Lqwb3VDo=Ra=86;)$M11Qlp?-yAWEX{lylyQ_!E*nF7MM1*QP{h+0bUQ*58p6B)zGeq|(^L(=V(W9!0)QrpUvvhA
z`qmpgnIi_?H>*Y8<$}=)^}vtl3wcaDu``Tgv)v`oQpMHVZbvki!CeZGOkv19wQ!}n
zV&@Mp12|b|U;;`&Ql3lR`F`2%NQ9x2fO2Y8dGU}#Hs&MEfw<|>zdHvLMjwZ{V=m>;
z*r);kN6u82PpDmEw$)4Ig`=HNFt)ZV(#_R35em)?ki+yp@o@u&)qf57USes>e>r|h
z_Iz|-r1NtF+ud-QB9S5cxg{xIFyHrx#|`Oz^d&6ld|;98o%q8IxkDaiYNMOY%1Z~
zc0@niLdqYB)Tq!HOLMYNgnjll8@({Qtw*sA1#ElDh>!YKh3)lOuekQ;kK?!x*x?5U
zNU#BU=&1r@Q8!ND-#VXZheh+9ut0n&i{#z;puKP#S->Xv?Z~1MvlKe-hg4Kfq|1DQIa-SO1^BNT!=x77>
zv(C0EE@Lr3`B-Uv+TDODQW&(rglTg92qvW)|yl_L1_jA!*piXP21E5riSi?zQPX^`$J;-Ht
zFd7yyM7f5ptpA9`+fbZ?bD(+-uLE;t`7}OZjkvF=*Ug&{zm03qLh9p1bQ|4nssB-M
zKdwciwt3Qpga%S!=0q|Ow>%^P&{pgu)pPpnQ}faZ6@pkNeSw~Q1w07go`?%zj7zt7
za{<`hV`biQ1pz?1=hGI}7D=LO@D`=L`-NWZZhsbSvs_=7H@0M+_}a}ql;6i~7C(!y
zqH%~&eSS$Q2WSXgI=zbEe!fJ87yn)#js>^yDLv7IWG#8dH|ibjNwsj=pbdDIr(>7SUkxZIngjQ;vS5EWH$+WSd;;0c
zcg5^hz=0#8{n-)5|JD>-bnfbSn>JJa#0X9E%FI6#J$~&9s1bm7-I$VV{2KWxa}+wE
zK^>=9-6;?p5Uo#pdvSjSC~s+$J^rQcY_q?`CCT$>4eV=p7rch)gVQB`8@uX+
zZoNN@vCnkr_E)FjaGAf;6m(*xTx;`wO;DkJ8s3=raouHHmCT7DNY0AL4Z4w1j-VNs
z!IMa|r1daIxBEJNJ_odC@>=~@uZORqgRpRw{2rRtJ#Uc9E%T)=9luv_%K#lB=)S{;pUre=%AbF
z?ixRMsNvMu>C8{G_@h}Q=GHJ-dMKupPF)6lz|x^*YESXB;W?M=KGjU)zF=Vg?U{rc
zXAzq7uQ}z$_K;&=Q~xi#dU(h~cPiiQLp&9u&H{~+7E0LT#1@n!kHw5cF=MWZI-!Mh
zz^Kx5CyJe&vG0r5R8jw-x>dEV=TVk%RjgSf7~CS>^t)y0s+%6u$hfECyhE~WXA?lQ
zk+08&`%{`T^Wqq*LeZp@d^
zzT^2pEj@g3+PDnlvwF@M*wx6v$>bh8!-Sogad|V}*}mDZX!l{W)BO0kxHrK(!~X#S
zoKR*8oF6HiJ}PManBaY9XhfPPqR^5%`>_`!WA!y)FAQSgF1}h@4;9j3HkdW|raJLZ
zH{V)RA02kDdQSE8yPD^D6#KlBd)n0kvW=Xl!__GgB2mouI&m*wU+V4FSt?D1=LQsqdy7JDnWCMM_c-}U!?Lzj&^0z(us}Pwb0X7?(
zDATsN7a?`}SiFwbuJ?y@qd_NqUt=y5U?EoUL`=#_Lz`9lAuZ+4t*4|cfL=W)dLsy
zD$ub8HzsT!ZOzaMjFET!4tw;Fu57O3rqX-5rWr`CHz*P!*#X=n!fGQ4fw*_w)=rRB
zuG!k{^11j<{_RPPn
zot|8b8NiuBOe$j5{PAgbecY>%(netN*kRSEzZ8WNYq7!8(D1E2Ss4EC1wmiW!bkpc
z@XlJLr_^Rr?EA1Hs~}0kiDfpntjYhK&O-fKhhX43vLw3C1s^Ma5!wd^5dBiUK?$|=
zZzrHRWwrtZpGi}?B^^eAcZDz6h4KfT(+Gd~EJgj`a7-qWtth!alu|)-A4V~p5Qe5g
z`l1aWqnSML&V(QeH`^-$h*RZN!K~OPsUc*J^A;CVp-X^h;?L!pmA2^*pW~lOLENF_
zxPr+B!`+9nk48niLXBY(Q%QCdYLsw$K>Au(WXeH_6$+&epw2;I`q*W$k%Xow29pg=
z1fuzOu&2_`5ScYUTwQk&)}Qa+oq9TlnSkKiBx;Z{ZX)l{D8~YrbMUVd-t?sY^AT_CM5HH)|TI4$2W#fU8)gy=x_k^cz%%G-&rE21Na$KS+r896FA6(^s?3jdY63G
zgE{3^iarnoIz9RoAW(*1JiOipcPtpq$rr$Rj(3mfCK&ShiFZFyLUJs%k-;p#XUxm>
z4F*zc0%QmwP;vnH$}JZyiJ(O_*@_ym^j9E;Vmb}#jY`?`WZ7>o70;95s+KG$*zwKO
zV`zMdoWmhZuyT9E67D})Ex$-6iwmX^!G`^~NjK}DCKxHZ6rcB0;+F~z;9qpaByYH3
zlr^tiT5a<&wHV!jQwqv3U&*H@yeVP5b)eB;2CGy(Nk{@*e2=q%)ouFGUBIkHBCjn8
zeqDVMa%%i^;phWXi!$NJ1S_ujI6ob)@agOpbPzg{bwVn%JIw>J!&j2MCzWCNoS90A9n|rZw2TqPy*nV0Ue9V_ylbMQz
zvB%Vgg=wbtVj!Nu+dn>eBmAO1eE~+2`=r?NAFD=b;-&=Q-|UUU9r~I~Po1R-?<#h{
z7402imox^lS~%JT-`%DKHLb#W&2|O`R_uU{7b6t(RNTY4GW@UR(sRXjL!N$&W8W_m
zJ=RLNoa=NZM_jz+nWmwSRu_j`3L=|wko;EU^=gpz9phh+p&yg*+xcJpy=bm9T-*EMj3YlAowfiS;ZI*fwPQB_Rceh0~x^e
zEhA&aZkeiNlUrEWynOmb^X+$lBQn81$j89Q?RxcJj!5+>ueuA02`0(SM`m2Rce@WW
zk7rI_E0F=W))8UCGR?exbh|oUzQNViD|PGoM|fNEfh_$G-mZ9G(G7SLQJtY5s4bcw
z#=+Qbd%9hdtNmF$#Zi|(B!{yP4!plB-R1D;_x10r=f3VvB`Cd+s&kI2cl0y$KdKoq
zyudQ5OC7s{%sde4)k;hiMaDWDVLuaa7U8H$U3M%!xjnOz!|H{rdZb%3gEco_KzYhB
z4}OjmJvyE*WOEBANo5=EdQ7Rle<0;pa2)^3V*+7f`d`yG_y6f1ZLOFC4wV1sA5#O<
z!YSl`pNh!|0#w7elH35(I8)rlq2`b{@)|Pxna7%){i5W>$c<2ng*D8G{IPQ5C5<5x
zHKkhU{M`@EIN$H5=cx_EUw*o|QU%YwS>1FF+bS;&JuT16!R=Oihols!rT5vw93wT;vCm-1af|Z&-`cF)
zn!M^2{{otfTiLtBn*?2&+fZ@CF=U24JUU#_Hc?d-)0XucZ53QL7DK0&6*;t(byxK6
zcEDpq=g?iBrXdv+J=4hz9sjaK#=PU1FTQ2ozT<@-sQ57WT|1+i`#VRQj9hNZP##Cw!afeV*8KopT`9_ijpw<1dsSUA$
zK0jf>v;==;7XAvP<0l`UXBD|bB#_<(XU2uz3&N8?P7~9V#~x3tR*igDRL!STD?Ft0
znFR>3TuD@u55|;4zt6Tp03ji*r)h*S)1e&+)7(1YY_GU&5)JodAlMXB)1+o0LXevy
z!LQ^eWVF}qSHiXWIwpf=lXggzDq3b3%<=9sFIn
zWJo?O9}Teow9fHGAj98DCX;G<=@*r`%M=85ne#I!
z?F2EED$YzWOK#r2&?aQogrCHR`$Bj`qttKDRmLbH?Z@FYT;o~<8Z;u-$?-yEf*`2p
zMwxn}O796rI*c@&9Q|aAorw1K`y`mxE*eG~^;dS?HA2E5f(_n{aWL(BOc{ztB>*(n
z^IpoSn5eQ}7fq(@iXrBY$;E)YMrAo@HeD32x*xm#Cb+rEkW*FdK+7SHnG@+)#1)T2
zWv}m#@v6M_A@V6Knl6n}YwLW83nt{FIP;>D;vs^@jb~|a{$t!w5#b}Yl>ePjAF=Qu
zo_~QkPy#H@xg_RMY~DdA=dyV4&!5)M&>xOUkRyjk(e+fdRDDkdlH#xw`|8%O<*xqW
z>`_p5d%y6l`<9}Dvpzcw%~<~wZ<1K}DifsUd7Ski8-orYAbI|jG(ygZBpD2%{ng`YK-sVvvFa{g3N@RaKsx2{BHFM-*D$i)k8s_U)i%de$}F1MdGGqv;u{Xim)@Vr@&QkE
zmuVQQC5o5K_ad)mOFoZ8EdUNkY&_hvY}{&q)!ZJyBVS&WOpHdPQ$_b>QhuIc=?L?~
zw7XHy%z?sLN4tB-!Ut5e=pJH}uOE!vlYkgY1ka~Sxg`}~0&ApFLdOy)4Fs}b1%uKB
zsWS-U1`1_QpxyOI|DJ1`MRt~!wilH97tofBosQaRQaW%+tbSWA7XS!MChYm@#v0Y#eZ)rmAPzyH#J6dA&P*XrJ%jQ3KiM9A4f)uin?L
zntQU@Pl^rwK#5+u4xW8UNDeFJo`#9>ZBASBOxepOIB
zeubaA&ugAQJZLV=e=Ho=D6GGBT|Tjwl&7rT|IjJn|8o6;c26(v_U>}~c&Fat<9*kX
z&{>fqUHR%dl}-kb9DGfPwPC}*JcIVN??w-F7uxhw<(G>w3iH(&c>Oe2J*Dimt33fh
zn^HAN3#m?^g91`Hefvsrp1Nt>h6>u81(dagsquVv1CYa&RgxV#B&94P+})HLLnSbY
zW08W1q+Sj62~>5Giz=?HWLwHvQJWj$v&su*NiU8MYE{F6J-1Q>9+z;vqRR;X%>oqj
zE;ieXm)Y&!jBJaLkL|0Vvs#JtAlT@dd%1#q;?vEz#dXSX*rNw>FsM0YvB^Bm@!=3^
zs+Usq;m|5O#4WUObxh}@$1^plG9&;ap~Vb?K_yIxYL`mwfe3XQDQ50eC+uFLfw8hO|35eQMKiMBj=^5H)f_wN_8976
zo>amvT3hZY2z_j!|MKzr`FK3-m_@_Min~XGSd~pqBm8UvW=6h88us*2M&A9BT~bYV
z_t?PC$qQ$WPsoupDSepcmS8p0ICST@hDQjsxN>RIE(-Fv#!Ki@j8+e4i{Q@BQ9=zH
zFKdKJA%IQ6LQQVv&B}r^0ZfT}!78T2YavC>*S*6$)l|HzOHXAA4|#TX0O)7&CStqA
z?WjX1pJibP)5vG%QT!;Q*fwLCGh^fyVEy&pbY>qunSG_OAfPp_!98!yfHjP)l)Y%Y
zpiQunIA2+MNswr!=g6f_@yH-!S5HTb!#tE=2H|Nk}5XQaa=I(=n(+B+5A=
zNpr?J@|lyAMEE!}J+B}IMvzp0&PHYQi?>Z5Anv}%APQgRat>fnv8lv+Hgd8h^lD(O
zx+(n&i3BuSmOhShIVTu6i}^$Z>H5zO^DjC@)1ape90H|u$CR;N`7#yXtP((lKkRw2
z{V4Y_91}uLfua&7Rj-Zumu=WjLXk{_@&IidlKp+ssg&h9REdvLX*d9&sR_lA?}T+1
z10KK3jq-O0*&=N0DKb6BDf>hQ&U!UwurG(9au^A>sJ#&k^d{iv0=;^k=m$|93((rz
zlZGcyk1kY`Y?>FlPoibkG#pP%PKK2rAzfxr()0RaDfep9T)dM(0Z?6T1lm1?lX9TR
zvi80WhN~x5|B$R`2(4#W$Fyv~TkpELd~dFF@4kwq`_w4je)=HSi;)zOIz>J6GtecD
zw?w!VElIfkjGpmcNQcE$P7m5*1YDzr7R3MHPAK30q{bQN9gpt
zf1ek-+uR|KFW{qfg}AqZI+#@D#{_<|0n1yc&s!E=*e
zW~e~|+r5x?zEWI`Y%jcLeruo>YGS)&`TGrInFA3Bcuk}RWW{sxKk$YKIjbrZc
z#TSQE14=MmCM%)!aiov%2hbt2#F|DAJdT@qyzZK58C
z^-Y=AoDa%Vj3g}irkBsJ;lwi~Ci
zW?(M%4gKYRp+=p9M44>xp&^>7^-l<#!wmIBQ-X5JZ3IfawZtx^>n(KavN(z(WvWT5
z-f}0BUB`;>NXFG_0nW%?XV47|Q>+6M56{LiB!)-%6u*RnS<*n~3x9Y;PW*x$M?G9;
zORm`{o>o_e8G_NuVuAQ#D;>xSZ`lg{pSX9$^r9fJmW1+`*@mBjZ&0vaFf4Zbx6d37;I@n8df_29dSzfJ}hxYne|L0IG`%)*k;JdF>GuJvVRE
zcLUw7TQH;?&>Xz*WG3tLV?{V+z-^O45N1=o;9
z%suhc?%t3M#_StvTYFcc(0yVdDs9ww~-`*6Yq*5O+!lOu`o$MPAY3&s`DwnQyEob>{H+kc_;R
z{3$x#0GnyVzJoGn8z-F^NCkA2WFz=g>rLTMX}1$hhK`9Q>tvmpNo8yIEBcuRW1Vy!
zyu)P|+zks}0V;XmcA@l;hR1twlnH^^PSz%?h;pABA^DTR{Qk5JM;(<{b;&jV8s9UG
zJt2J~aTN%{Ye;RFglHCPiyC&y%f=99Crx;gN7>`)0p_G4QE|kOob;E*VGYAaqS07p_D~V}Yc{Vq-JI
zQDBh$Q;cLgX=#mw0Vp(@TwDf<(z9bKD;c{{5%5GnC1=}Fb$7;;>O|`*O^G{z22MnPeu&K`A=@h|2*<>3frCE4-zeZe^As*oI{M3K1fr`k%0GDLc}F7s4F4c^pT1QH8Y`X8Rps@{kwc-nJQ<$MHcAWmMyqqYj=ZOb7nV8`-50B9
zO*!bGWf!(lgcdL-Ha00T>X*Iw00fUs0XhtYE7*(QSB<-4M1L4fR6wdNUsZ|xJRI)t
z=_;rx;1wI;g_bZx4ELr%+D?&jCj125{fl|kOy0m7+Uiv}bPG2S$qXqWp(5?QO^sVZ
zeM_U(Z%CMLAwf4Z@q?xZ!Wbp}bKRhSUapt-|2EV8$wdHI+zgf!D#N=W0Lq>c7iS{;
zOtiq4zInh+W*;mX2pl3(M>7<=NP;6Z?lBtnSX1fKCW+G)xUw$iW#hW4gmWUQQko&D
zC-0lM)g1Ul4ga1nLvjEqT&%^O_^PByEp0x!JM-TXaoV7G{e8z9CPq@ztlizmioQ&_
zM}oCvQJpZY43lc|CoUS_2b}i*xRyW?*|iG=f6p38y;M1j8XcQ8MpZ53K@Tlw5^f@I
zkoM?Sqs2zBvUq>Iu7d@7)T)P_9E|AME`C)+=fGr5f7_*XJIa8oQ%dnrOTxIx
z%Jl2~!zd1BE!(Op}*hZ_#+o188$G#>=)pWO&Ne
zM%@l!vq8F3Hz-zF&6s0pLPNy?yLb%*YQ?dmCPZA%V_VIFQqLE42StJVGcsVTWyFir
z4uv+nRrkj$*l`~oI!U&x+62ZPU5YX2A<~dJW!XKyfr!M1iRxyP6#lEXiaNuXEPC(L
z2`9$G;N^u9#uNc1fFqd+M_%s+X^V|E8JW)4vtC8ki7N}|hJ>HTadWy9$$RXO!+D0-
zNhrCKj>@{&wG2*sU71fD`Z_g@e9WZ7?@9FVufpKp7!}Y~k`8mcU?GO&8OMh0oUpyR
zmV-{6+A5L$=Q}VtC73_y3d2C>ZU)J%rsm?k+eGl-B|=Y!0F}55#vA)@S5{4Q6T>7T
zCI|$`ilt+Rhl80nYgooA%3PqY?5$Z)n+#rNjS51$?{mQ=IVghU{Qe1NUY@-{bEaP)
zAfUmR#uG6KMlLvimlySbE5mDSSB(G$n)JU=YCqBq#cI1V(GYedeVP+B0%!t%y`9sfr;OBO!-#HU+x~prKyi6I}`c{w6o5
zu1%V#sZ~J;*S7cwM8drpL3@fYpma^p%W2Y*U(;po|K4&UkZ?7vI&b;G|T6vFqgNy1=S1#o~U6G;J|OP*OiNTwZX#Q}UR-2QZu
z$5;$Wu7ijhxJ(>5bX&NYtKS{_q1u1Dpg-IY+E;hKt5amin##9LO>Ig1)il3l*dC2y
zE0Zi-=gNR8x;j3tePb>cqpohzE>c>0Twwv=jCbaqJV*OI{C<77(`a!MUo0a{TeggLrYia9h~JUN14y5RfJIH6_eSUoZLdBRVByXBed0z!dKR)64EgmAPkH_AQ0+F+qv
zu2520Za(o49!*>_B0nsyw&6N;A#ODr^F>cz^bXX9r7MZPXd9L`-l-ocrecl)MS-
z;!6Z)OAKhGr^wIi)k}&E(>YPViaauG$DxGeS&&~UB5y=$@A!j3xmfr
zH|G*nmsEbuNquRY4*EEXVd^%6Y2K7*8s`_(=zOnZaq_%SE3RZi0o!dz6J8vFi3pYv
zDo9RnP!lG@U4NcqyUuP6$Z^Mb3HT~#O^=1mw&v}{0$*1a6*$NMssvejzm>yL;Iq%Y
zsH98uQJLA{uuz!J*z+x*e6)nmF26ds(g4JrAKbt^aJNC7jCp-!#QLM|k}m@E-kG3W
zB0^Svc7IZ^Re*^Db-7*w&?;Xwqr*(qRQ_8#!1ldhM&zy(X7y?AP-DBlj=u|rE;F9r
z@L@G!TSVxlmwCOsdcs)z(BCKvcp%o?91c*@EC-8i!mo0dBB2A|syHUhUn{(~8#(8t
zW$YaT$8*j^iN>eYofe^J2W536$%p4-Lls~%K$g+|LnmB66LLIV;VW03nK8h$=tJeI{v8WQPSfv@eo0mF;Y(cshM{f)t|7T*M96mNqx!dHh?2G0puCZI=lCIs((INhSfDh$hXVIb_HHq(Zb^c+crH
zul5DDzpXeZ81rG)tfkc7n=B1_^h{n!mn4N=()Nx
zv_Pr)kyKX1<+k}r?Tk;gJX;GOq>83=MMUw}7pi5?c>?=y{
zD;mRXe?q48D~e9la${*-o);2YghkrU-PiXe*8MFnm_s#^Kdey(U58Y4zc?+5a?=f5
zuA=qsaY|XA9d?L7+kgzi?CXA$NiMtyfNVTwj_(&%CbWvMA%_b5voH$m*{0L`p8q5w
zQv?sNPE&K_L~o>s&I8AKZakM3e;oQZ>BkRkoqT
zfI1fVtThoeE`(d^D6rq8f-b%`+dhr_XoYvPkAd^p@OeO-BJ`L|Ju`9pJISIk&?eKb
z?4F2RI9>}W#ra6ZB&yVGAwed-P*`t@Ca
zkOchc!;Uy~Zcxo72@nJwCVJ6GezP%_7wHHS
zeqzH&^b4X9AmRdMh8Fz<({UOK%HW}m>~zTbQLXqvD+((}UKIltr90@&h^p&=PjbTc
zbId}Ez9+U@4a6l;y(_UQI0$|G0GtWX1uEF}ku}Ut%^X&vb$1hbCoEUsA}EEXiD=8W
z_6dirqA1ML3m=z0l_fempy5QoVU6Ku)nH=$C%Ifts}R+zADC&WCk}U3vXNQ}`S(7g
za@%Mj-fxF9^}l+ga6acB1`Pcx>5zCAV{$-`5GCt&O+~@clhT~Y-wGG}MXn{=MEVzE
z4n>%f7<5B-Wz0MiduKkzB@Jcaw(5DIzU)IK==-QC*&>6FU+B(6aj
z^DtsVX}r4jQE%|WB&VwFx$qnzNW$#s!-i-N5t-#A8LI5G
z<}HiOL|XC4#09D!)^48nlTF~v*zr1m(@Z;w4o#y13BhhC`j^NBlmHTfy}(W9^C5j&
z`(@{tB&&fdI(U)HtjIQ?GS{*DZ52XsCUhr7`WE+#KaDmTb^<+=^IL!!$ai}0(qaYaVyzB9tE=okHajcorH~#96Aq3@Qh4JY2@SC%~UG=`IBck
z0=LjLn05VJ2;Eqt`8O432E2(!zpdXRu|zj>zeCdQ!dW03#ea$J4PR0
z1DDqW3S?1EiLf;JYcwQMf0RW?`&7iRW6UQ`A55F$7E+lxv?`3X#=WE)q)L9pj{!{9
zY$r40uk3IJA~WO1`sLlD(k?UqnD1OYyWInWP2{r+jz1G=DF8CQ?)NAqAX@9;_7cUI
z2dKtU1BxQ?%s|`Fj-&{b$Sm!GzwJZn1CS&l8>k2*`8CV+pX_6iy!`5tQCTg1yiFSpn+F1fMxlrGiPVn{c8Ltdi(v$OSa!@O9@5ixmnt
z2}Oc^{Ivyv8%$*Q$G7q(F6>4$wcw%)@x+T0jTk^JrRlBH0{>0GsXV5**eXsx~SqF$z81Yn{hIyD61H~290yyDNg;zh)25^
zw;jwx<>+xhx%~cxS^+>I4!OgYwR&gHjzWjAC0tuJrKuGYtUI+h({5heZXPHfQe3y2
z&)CVc1!oK2#DcChh+T-jg=m$kHdl6E!O-|-T7rYml5pfV*JNHAFK(ONE;pA}@(Wg9
z5VIUGN^#X)KJH~J@xt38k-gN(awQ}XHzY#4nQG9qSy3zin*|WLg|gWDem%5AUD3Vh
zQ^_bS#-2;KQYoS&@RFV4NbJx2%^F}S@);G=c1K6*v`YXM*aG*t<$yeQ8rv7t>a(XRv}PRB-bE+p-^PNbAFqL!s_4A0ki?~l
ziK@y=bSghOoqv9xr{~Rq5c7?s4%@cVvWQ5IQ6p$HzET)&!GCNU-x5vCYC_$V+HJ8v
zS(_~d7Ynx_#|-jMlIELo#nl;~45}|gUxl%e#6hA#S9%pJH3k;1inJn20->4NJ*n2w
z{N_$pF>lnZ*yf*w4OnH!$M45YCYalr(?S-2^APZQB4Pkqy>JUBCI)tSBabWdOfaVU
z`eD4Z_n9~qc9M3CGP98oCJ`Shk{M&{Umm#=yWLNinNMHfpNlgeVVRe8*cW_>UHn@z
z4nk}ED2eZ{2EQC}lw5TNm%x=3r+E$kRFUOTI@}-TDDI=L4eGD09$HH@)2phi)rxlZ
zn5bxpFGavI5!qGD&QbpFXjF9tqm)Ib1(##3v8?#2O5TNu#}Ws|Jq9j#sNpXyy6Yrs
znu6B|Rq=>ty|yNy3G0>8X?z<2z6JIUlqsa+G+0&oXPw#KoNKG0#&xR|#&@QEO%W1n
z&|7Ps9F^%scmsW_w$b!>_Ji$-h`Q$+g`84CV`$}J)d2K5hxE68XY6aNk8_&tZ=)_qZ{LJ+tEQpH
z9Fy%X`0fDvCg>3~Nl=S-n>k&P?r@s(COD-T;5puH@25$-hYg*e18H~tG~ag@iw_2J
z*ETdg0A6~!>n(pbbIpS|?tVCN*!e?+o(GYF#aT+Wx
z8YXW8_+apG$d$Z22lPFKzYf$*AoPi<)&c=s(vZd5G8)L2YD5hzx2}d8k010JGb3N
zIlvCC5c~|}IQWc`yuTvMfBB=-D0<78@OBqG{-`n8TOHbIsoR4%D4htZd?+3OE*iU>
z7`q7m4q+NDRBIwiYtie82x7D?*EkLAyM?VM1aEJj!;5}Z<3HhPT7dmOsQ}pN=|APq^odf7
zd!Hso)Z1-d1z7?83$&3>529^Mw>2IwER`|!^~5wHv=cG(&0bB@H12P}}AaWepQupFwf0zu1XWXLcui&+ie0V5M315n7a6AR`(*0bA
zY%^kB$c>7gXr$K4Yywx{PQ`kJW%Szs|63J6Ur|qMCeyhIbW}&E-4A_a!S}1jhn-g!
z)xK!b-VWE5Qq^yPT4`CE3~lK4B(KTL@>PLq7^H6$1VF?ev9-lbo%hc}Lw2ij?S?feMCjV;;=EqjbXmk;CFNh~6A0Li5L5^#rp4=54SAII?#4;lG$8-W&J#kRTHps#_JgMkn0=*2t^NPu5fYLGU%tfL>g
zV3F9kAsn-~YvZGisF{ZH9MNe7POOtL6Gp$jmy(NC=rmkP
zoANW^q~s&gzfL>wU#m1c?F}Y`(Zf+U
z09iB2?38iJPN;Qx;Kj;w^$*+c9@OK>Li{DE&Y*6zk%j0cw(`Jd9_^T{o8Wo}>vIUE
z7Mt^i#?(>t)=0zS??+)K}hqq}<>Bqp(&0RehcwPlU`1ZSX67u^hirR?*e
zedt`MaV|Qe)6P_PAy&~1o2z}gHip~_06iU)+})5`?K`xaJP(9w>02wAcZp+%0p}$1
z(!l1Nz^xKdj6pwbbb9pP;Y27@AUsBs>huVuUov+0QE@W
z1W-K$9@1MqL0}@J4k=?lNspe;50@*}lLU@1!q9180=RW;Aro&Ui+IgCEJGEXJ4sN2
zUcaz`PISCo&=r@VgT*6)%m&En9npYC&%@0s&r3=7PhJkI}cxlg6g~?0KILY6nO~9Np)HSY2fK(L-{6qDv`@JE{v0Pik6XA
ze*B)~7?-p<-(Nj~!NQ}S4|QkhOg+YpWH>0+0*!Ok{@V&mVZcZrC^d^}YpzxsCU2RU
zOy&#w98QvqxUM7zku9%Mx|wiH!JhGNEtB=7cjPg*sAstNLE-{-GxOsq`XO^1S(62)13ITJwk#*wYs8D$px0%@yhlnZMzs#S
zGtd^a3X}||-`*EeXi(BE06%qRdu%fLf&~PT
zxKh*MdiH7|=VxB6EnJ>N5W!2csK;uwVG}ezrd*(qzJUFM)bn&0^DMgsqvXiWX6V4#
zw&Am{yx{9T0=nccW+_s}=w9a8Ku44QESd*7+A7Y}1oW81%m6tQKr8bm4JyYv7SUXQ
zvit?v11!lly7CxGrtv=(b84?pJv%rtW1pSCWu@a95-i@1PIVCEq*axeH01+ioXywC
zkR1~4k1adnq)Rm`9+-%PmzRt}DY)y}6E{XMR^**s)~>#GV$RUFqEz~7*uRuWvI83Y8mYcx_3wDKSeJZCKc`P`
z=3%(@eVmA9&R0oslh^Xez5M8DV4qTc*DmmIF)LgaBNM!?re7L*XI#6npzSKxQ7L()
zXH(un6FPHd07*6^;L?}$t=>*5)jRh$`QR?9vMf1B)!K;YRn)q_gO!
z%;M%va`@56S@mB^!J?cu5ff7u}=Cbn;f)bkxFP`KA}9Ex^8c{)j_s8|
zS+s8Kik%(XRz(%twrx8*w(W{jv2EM7Dz=^6>h60!o%0{oGuOlz<9!C!-i?uv=!jZR
z@xbTB<)?&3$#h`%UUY4*EHrm9Fg&x(0Re*`e|9BuvT76AUA2!M8t#f>HNS)Q9gRL1
zX~IP>ZF18Q;i1J%G85Bz`{?rL!S=FTyArW9?EJHjaY3&vuv|$8i`0qWWs$A+7QWtM
z4#ug1my}MUpGlA3f2!Zvt#=_I!>tb@XQvlc+b0ExCv|Zk
z{Lza0Xg=CREO?Jf+b;$VcCxS?PR1A$p5aA@=x40hudnl`{j>F$m_<2KMix8DP-`9n
z8SYT))^7U|MI#YI8*e|0P0V3&Csy{VUdPc!7J5dCN@jWipln3?k}6Nmo{lmGg7}gN
z_Y^KYuDYpld+>2(N#F!L`O%>hx&7o!%eS5cLM%S^+Mi)U21qD*}wy(u^@-aB@o8GVul;<#veAYAFo3C;x>l05O_b*
zUv4!mweTEnz+JsNWHb#&?GmA2HYEJB+Vz4>lM=tA?2?PW{!84y8N`P;2eh2SvC9vEW%{WoJq8LZ!3`#|WFvLr|msK@{6f@4vO&4(UO-IL~2;%eAh=UMEw
znw}jGhAf!>ui>8jTMdDV>ooQw1|pFfP9;a?lI-Tk1Y|;*(J~FV`@HUX08&eZCv~3^
zO5lPWdOjWn0a=JD`JDyu@Pw=hTpud*=R;f&wMLz>c(oxlEX~B4iT%gAm)+>2N(8bGYZCm=JJjB
z6B@{+8G)d8*#7Dj!*B*x;RmI`nh_>E@I2sf*zL?a9HxF}tIO^*^3Do&6{WZ<7R6eG
zBjGu3OSTB^se@ftD*WxK+GZ}oaw{i&dF!h&x#tt@o@URj-a^#bOCLY9G-D3(y)`Z}
zhVGTuqc(Xhcx*fOv){}$j?>!5sexcq%7*g_vsf5vMlfg08i5lXQN!Ws%BdL*c>V`>
z+7P-)YD3fZ4_++JzY|1qtWh%>^d{F!%s&9S!-=>^Z<+L!pX@VjnHLoOX0GnL8w+hD
z6kX8)t`|aT4417&(j3bS0ru}{5z-xq(c^#ted9`ggBQN8kyW_wRxCWmp2^kw+dris
zi!owh$92z7LyESBKip~}3ssZFf$t|Ta_>%Y_wv4_r*84yhk7%06bt77u*-xcRaS#!
zo5$H{wbxG&GAmMKr)PWWw{~llD*x(}DyT@kbcIEP#rNb%J+30xZD4#qXWVY;kn@Cc~=Q-Oki
z%Z*j~hJU7jr2n|PS3+N;7nT;!P-j@U^CJK9J$2Ocb8$}}NBeIl7hJDgzL$2(Me*=U
z=MGj4Kdy;_^qU9oZ0?Ig;z+Q?$_NBqw5~PZx1!_Y_plydO%E9_^?D-!YZp{+bM5W<
zatNtkTes)G8+1_CG{t{&F*rLDOBylJKLB7M7rNhd&98N=dCPx@AA_Q;u#^(g%LA(U
z=CfE5=~|mIj+2c&RMUDc+CAMy#(tzt4qMWOUbb6PxBwUhiIo+J0-TDglAV
z0^yD#=8d+3VoSJWi}5Cg1RsUkZZhSz1}5i{~;15UWA+}REd%0Y9s
z@!`W0O7iR$!AAmfOw~9*u@-PcXFS$-@usleUA1nx;B7;->P-^E*d9?!KZ%Ry#O`I|
zII@_O@UWa9Bu=V=>)=sw`tSSg9ACrFHPf|nSIrwE3R5eX11p*(9Na*M)0U<$!Vyey
zB5mi7xUl!2vbFXt>+PTHsD9Xfc39_1SFS^6vM25pL4+PJ%@D63E5SQ
zI;>}0GA+gd1SF{S@**5l3yZgX2j+r0UB}S=tSe;vxFF4%wta*UVwS;^S%+A-plm6N
zxF9S%L!W{29p748A>SiY-wR*_&i$qJy8*tP43gnFM)8*s24x)CZIo;0oq{~pM$J>g
zMh#`PXS@l1VyRGMrXdh&OHn{F$G#zf4q#CSJ!S5W8^rTTD9367YiREkny-{84A2IB&{5YRZ`s&x|r6s^$EM}{#l(ikOZhuXRHS2M2Jf1Xbv$-zB{T;MZ_2QpxT)q5_XAa6!D%c0(a72i{8(
z{UB?L5EiyEaDg~t9)hDJYEn~Kru2OUkJHR@h^&0{4vpJ5gpy?wuX>etM6Ey+s&2&Y
zOrK}DC9jnnE$_941amG?2);82v6W%F?4Z=w_U=9jL6{>EOa`ZiE~qrr2`iB$`&N}g<>&|%%@~b82}lalXnh81$^;kv_|BeE
z>^$u39agzQ>`^KbKSRsZ8>{Vg8bCuLS;V0Hk&TzA(d>;3m%zRIV>|@(0v1?*Oh6!6
zJEVE#bqUd?=1;fXFwBA#K=#!SDb3Md2}nsgZSNW#2FzqVNK#Etc#O30V#o`P=CG3V
zrExc~!?(KWOr>jP(_buhf3o$wK(gi5?fkd!gR%Y3oDIgw`Ja6#H4tYlc7x+)OP{!r
z$^aZ?Ko~lO1c}-!Pqgl&&H~RyImmj!y14m7$^|p4=K3a{ZvD*6);_D`BuZ3qpv|0%
zb$CuTxUfcSRY0Hv!1tN>S1}+m+HCHSgak(UC72WyWmqBimP*=LL=#uHy0F5LJ^$6C
zBUgdHJ9qGg_FQrj8;D;K;#u$5hRIp)c%D&i&Am{;^W2VlB*C9b%k3U^beQFIrH8#6
zWKW>NSR6Df6_v7aE^%xeSL#LxusoA2l*}#~dl~w>bjV3GluIWWN#$N(!}|?yFn8a9
zDy#ezsX1TJO|PE&ri7ibgMBuC89pvx-e5~_mU|BsoQkjC3mjNrq&}xk-$x*q1l7@<
z4uIYN1%^lVXSWFQms)JP0o;xxgcIqs)AMDBVOg=C1Wf>Ga&vet4@~Mqm8^oCg$U|k
z;$(pDF$#uaXk@k{e5j%nOI}Q3Gi
zk8ppB`pCuzf%(Rt09nGqmHlns>0#v0qmh;g5g>vEOuR3?;ql}QBlhS9#U=n7dw)5w
zx-gnF@wiq;R3Nt7J%FGb7Em?{iXn!>@Fr#oP@E--2O8yO$wOzem5_;37uJFx^d^UN
z>Pp>w35zPma$SAwo`d=T$UES5G9zjb&szY+=OJ!fxy47QeO5l!^FFMPo4~FdZBwC5ou#|Sn-`taa0yECEl99Qe
z?xmkOR8w?lE;)-a#b^2Rtf#$YDjpv{=v1HtffGx046MEBqIG4QqK^9IALvM~#WKI+U3m`@
z@Kczmnrt3il%+>M-xP)RbVB(U`NRD?u5yRNOB~pSu11Uy+&~J)q9hp^uZG@7rlUKY
zfb*TNPV;nFOGCcLf9=ub>FduSh(7i7D`EvrHBqK*DcIf%R
zg7Fr*@APY3?U>sKPL|}MFn?eJ!jLK(dn_p7+~+F%rdQj}yv(XEWGvP&QRhO!@JoT%pgA+YKR_!m$v0d!VeXuesngobq
z@udIma+l3VvQhq7(+0|wqjH*ai&fDd^O7k2cK52yg-IeBB;zci1_eCb>M6FtPP(bG
z4!2OyEd_Zn_*$kZCGYKNAw+SWXrQpZ?D|~nVytH`ObSWnAUDG>mK5F{88oj@76vDg
z8=R4wDE(GR`DFHpjcrzayJObiWo2;IP6WG6xKCbY=U4a{`?8t(y+(=1R5f9<5B2&Cx!Dj>zPLGZ|zfk^I~QAcfI1H
zr~xr|&cdk^vpO|rSWNfBE+DLfcZ#t8`(bm!Gp0~Jysw|*T+RA+4SZKhla-Hs7@-a>
z?han{KR&G)!u?+upJ!EZ-N`1^5Js+}H#_HHiQRy}^79p^62gsJn-nqd8Rx@L+TOfP
z4SqGs{jF)PD_=QWH=M)1VAz}0X@`{-2p;fBq)p2sBw-w~Dc>-(HBviHnp_?R26VcPpQ!UK42fSF+KOU>l3
z23FpO0M=Q*imk0P^}~val^OzP#CG|!pv;zB7u!es5$I{Ih9vXsoqk*GHMP}~ev1Bd
z5Y*YFt^*kffcY4XH?aG?@A(;?mPkAz;yJ|okag|i7Gm;?81$PQ?fm%|ksb|tfdlao
z@?fv=vb26JQuS3W{rg0|+I)kXKvEGkxzV1^i%{g4>Kv1GuSdA4&?hW^U9R>ae3N&QLo`IlUQMcI-1?ibzg@F@
zL;UMKPsd)RD((xq@Ay&e3@vP_fIGvX$=&7npD`4N6afQUXIsX)t!K*u1@QT={!=~^
zqZ9s56=g~XiUH2W$@agL=yUsppXlG+JtX@i2~t@R2fPU{y}te~?8rh%3@{3Z#dg*b
zNmP!_DLr2V0*Ms1>dPhIS4;eKu5uR#JDzW=xyDm;BVMli={q}&a=G1u=^7}JLA?zO
zC74gJa7@i$j6!zE+E{4kg}SCoC-*};l!$i#>qjc{W!%H93bd10po=+c1Yl;6a>RWX
zX+r)N{{V3$?A(Q@q$fv-q*5B2<`wfo#`W9NebX8`Gc6S8*YJ2JOFaI-Sc0(6s{=wQ
zfvB^TX7jXky+L4XjRlS7{sP|%wWi%(1_J{n{XgHnVhzimIfjbCAg8$EO`h9S!9({z
ztErt*&$TRPQYcb%U>mcKQde6B2TXv#!@y-D-*~E}@!vl0ffB}J-Jz4cJ7l-5P(#TE
zoQ;wW(kzuPJ|2xna|S_vSZ@?Ztj{E|NyebX>rcw1p1n`0Z?Uw1Wz#=~GW8c!+=nxq
z_QRT8RH=lo$8q}n&#AC29Wvpc_Ld?Hl6uqAZaNVbKPq9+KFFXDK2tPEa?T<~
z1&9a$xC+0jaH7UwIF!G6DDwhC1WJ@?wH`=8%2|KR%bJ4$q8dd-Au~ySJesQ#55X&A
zkl>G!&$ohaQ9R|p|0h5k1}kIK-|
z>*I)Efm1PwMS%*@khJQjcv=GoxyymyI{TL!cvDy+om(tDAm!FY(Dv&bIW+_|ntiME
zjT0Obp`Y6p5`LK+{9L>mlg{|Z{pj>W%@V8s>HE|9PYghw=PG}x70atZF&X!h_Ge
    c2t03#C$AYUd0|m9qMDj-@kkoqSrw1EEt9OX=T!kk!7i5ai
zD}}V{;hq#OJma-g=G&F65<=~qqRxXygLICE%{qhEN?1?hQry&O5(b}=bvjK2ha0PtFv3-%A^7BH!VyETE4D4hLO#e%2el-^xs`j%6;&Xs;aajWGpZo!-q`Y8
zp8}~+a(9L8bjHrr3o2T>4^^F>X7udny_-LalofQz!7Rhgdi*f@hEL~x_`UI<1cY@>~)xTGZ6fGZSu&B
zBo(sbSy#*NRaESn{_sM+^^A2ss069om4u-Z;_uDI>Y(m>9hJg7?p{uzhbz
zj$)xp^}r)l+*9H2!Vx`{^bXFpp75)~v#K9SD_q=K>eoodT+id4`AXEs~&7F@0wSU>5u^l^zHFZ4bw#}nNw!^J=CGf9-qf5(Y0;i@w1h?h(
zTu8{naYtN!B
zP3T&b`EfKx(3n01!8k^eO3-E3l~nBXGOf+-6C%tzRGHWaHlnVcx1z3)F25l^P`?_$
z>Tu_`SUS3+gb;X$IK>pvPXmg8hc_!jV?CH4!AU6QXkaYtY57v1=wO^I|BE@h)I9$W
z+xk1Vr+@WL5XwmQdc@yPqps-$%4Em@0dJlp&z8cBR+>Px)baam>o&Ij&_feYx&{7f
za4(VI-s7E^9N4!+gx&cQw$+!J|Mkce9kG9wae)mtni1L%@b0OGP|x;U$b}zm9^`zj
zzEmwz%nw_b>GfjwaUi)XehMVIS@8d2yE%{hxP#LOGi0IqV3|OjXn%+zBZ%%_ln=jN
z8BoM6)@u;yE+CmkJm4ei9ZWdfF?P3|L#;YgnowH~%}Z*PwyJ5&L|Kx`FTID3n(^LO
zud?B3P7-TaY(qCAPAy%`;;4uWVRU6hNwSjd(iE=Bpgk%`?-NQlT>_>{$-4PN1q5L*
z&2GbZlWg6I?ATb<*)*<=pnvuAD{}^}|9(#9MMq=Ax(tF``7HipvZ>B8`Nq7|tM
zXXDmy?S!X+k?w3PvP`@PwImy5+F=}M%y;y
z7I&fa;rni|5Ld7&f6$qR@>~bJSo8rVzb1Vbz1oZRGzo05oC9q)8`}MnGw(jv)q23&5W
za@uigfsdd9wk;xl`BA>#g4!P(cObkK`v=g4O}y3xu}3-!NK21KPcn`tPDHiuL+&3b
z8qm`^Z{Aw7_W>7bwwytAD)`65`)?Ws%}IFO9>Sk;`E9ti%}>RvdO4%n)BtR^3aUV=Tzyp@`r^gldFRFO
ze)fGI?ZSw2>Y;jh_%tdm&)DiZ;C-}$%~#-FM>v>$4+qA4&CXMGlMK7@(xJ&O+ZOIX
zIj+yjyaPoRSxNj7Gxgf~r+CB8eKSNee7g?P$h#R~pbVdX5Yb2r48UlElr)%OW2EY4
z3xYRZoB&>s3g=n561_|EDbZAdBz&G@t1|9sJR1P9@eykExp%u504=hb*2Ss7m1;JW
z9AY|K7^IH?Vu@PZ&}}2lu>@DL?uAh|RT?J&Bp|7hC9$>6RrilU1BRi?HLF`QS+Wd7
znD+HJ8furWdy6G0J^Mf8Xu;c!W!i=6${@q?5F82W0}`cj5@&z$T(N0sDoK3BRR|w!LWc!
zQ()RM>o{o-T{9^SEVLr#yC^cMkAKcpf2c0892guR5_rPiLX9&arWFy%o#olkp2?G0+A7g$^B|NrRQc#G`#UuL4~}^3uU3v(^j44iicJP&(+sxoMSe*0F@K7a%+4I=~9ErTN-Ulqm9>O0GW~Yi6V|#?^hu<;&Fr
zGsjjO1G7f37~rDn=gO$dqE+y2tl|^RIknlrklW&_s?%wQRVtA;yJJv$5E7kLmBH9D
ziId*kbVhMW+)11Xrdot3TLllm~5m@X^`$le8uYUi2r>~$aX*N(dy~=IKq$|KiA@?)Lp5f0NXWg$W|7G|>%X#a^;*&(m{J{Legmsn7r~{*
z6wyr#yI79aj$UL>@p|`NH51$GIzp&D6(Sg6T+V}~`
zMZtr=f`_(=O6O~+vqNak46JAqE+2I3Xz_-)(cb+Ij{28(Azu~)Hk$gCK`G`b$qj<>
zNp>7%33#ITWa58CD>o@=`_DLup4R#E-|xoE`M=+7N@FW-V;BYatS?ALCDccmNhqB_
z7azuE0aj{jP^A9Q2OG`jD8CL6MeQh>8Q0&*$RUpVduk)yywM%r(6G7o`Z1-D-*PW>
zJz`EQ@a^<@F@&UegdtcL!MXidg#{&wsGX1HnJ^ke7uhA73F(;VwE}(J1ne4D)?aF7
z?g<4-u2=5&_vlv{qr9#6vr_v=PS`qIyvw<@mzj$Gb0!che$q^T4sW7VY)r()sj3EP
zKhqtZY@bVNtROpYEH<-#rp^HVV(zS;YgPGmd|8(GBtooFJadrz*w1#UNoH%RcCTGF!8@eh~U
z;T~&z%VL?=L(vVSWc}_g!T34D_xy48Jo6jPF|3Aqc5u-|gI5HC8pf3u-ZeIV7=H?Q
zz5UR>pZ1P98R);%XGKrMw&Km6j)pmrKByRPVuJW4fx`i#oAO#nGd*&X{eq1xc_KZK
zs^M!m#;er6e2-l-t#RRjE^@wnPMgEX@|Ok#$-X}q
zebS(NIe;tONkQ$!4N0BTS9N}#%KaONnU3#B#XywL{)5W3wy~Ws5WXh9ulHgy0Tuf+
zG`G2I)E<=P9YOU6_+Ca1XME4c3j#O_BODKX{V$v0_FQP50fQ)qpsPl@^j`zsmoLql
z_W2gXwp&i-CLJvzg=Q~oU?q@+?;Dhl3`V?G={|lRS;mI~jrM^G
zX$w5}(scAv;;L!s+mC9r(peh5yXCP6B1&y6eHGH5YH5507~N})59N+$ik7C`m(-l0
zJY^c$x*0*V=#UKP!*__7l9ELLMJqUElDm02SS7{id}M4mnJmlGzBMcXbR*E^!XjQo
z7L*y$%SoCTUl;^@zu<>~J#19q&oJSJfn@zxmdH`BaFS|3;!${!`rvNhI5R~xTwVc!
zn5bK1YT^^2YZB51HG~W4YEBs(wH0$1fp#0+mJ1rBC2P(46MUZHoTNi_(oE9%S^L#rPp-N-?EJhsg&mI*TCN7y`@dX)Bah@Mj^(>Kt
zU0<#aWdHc<=4=%D{I0AzokFvAYz^M6;@YLfZ7U(6_&SeK63+bWnvBW<9US|a)5mt^
z<{|F4{?~n`#NxY{mK&*IA&5;fk7*ryQNp)9mdiTGZHR!lmoU=qo=k?<^zC^
zffBAdqPso3;Fa8
zBln;&kZ2A$?gD!MdE1au;oY#;<-K}<UI@kY`b<7xb(Ia)uVj%bln<3
zXbWBS9jS(H!4(Lk{-aF*Mt7i{?7prD
zarA{QC#PWN5DB6z`qS(*3aX+Rcfi=cn9xKzQwUJ7w}*CiL=>N2
zWYRh-Qkx)Hz2YJ;Nob$D{#?hs+5Z?s%5FsgFt2wtYoWp$>9HqTd%p7k%eC=Ke3CC4
zJA^%DgH}4=-*a<7H-C-}p_g?_UexO_n$R9)COke@Piw%^=_=v+!5Z$Zp(SH?GvD|0
zjOCM{o9&WIN;m8}snWMER->wW9n!y(@HQEeGSsUGA^=DCzaM4>f
zt8_np5(H{{Adb~)*m3(lXvqB_Z>u!ipE_S8SU_B&6`dvTNpuc0PDE3IY>bYfxY?FG
z4xH`hgYlnZ3$~ng%gl~{)bdsTXYs%C(i;S4zO6mD--KByiyA3uP|ojC{d6|^4d~bw
z14R2kLQ2zNV(OjEGyK2yAV?M83OjTmUv8mN)1CHd-?Bvan-CHE#RTfV-9D7jE=2ay
zrhIkT>x(dePHjH|vslYp7kT*l7fL<70$kHfQ8OPyllw<~BhFM>t{VK|3NAZ9F|41=
z$?b1|BXQ5f(yw&}}*4eDMk*Czv>BzqMNcp&fJuf^cemy@tQ
za5K}pXj2}H+Rw0T)NA(x?MFPcPz!s!A`46+nGZViW#wJ-E!|)i+ORA6N1g&B52a%X+4v2|+Qg
zYkcMxZ)QXtkV5Q>;dk+fn2*f0fdWC|#e%_Wv`V?kCmR@os4Y;xemb|Mvtn-~0#k5sw_e|r}T2pGH#1w*%(Y=Z-@8{3{vXQX1;6+QF
z$V@LUIz7%2mNq{YUOxM6{1Pm2DV)mTH!-ve8}2TjqkgMn{jO$mLm;HrUExqm2*8JL
z!^W#VAU5J2wlenleJGk8@p^d{2mWXU%vxF(qx-8t&{^K|J>nws%6&)2^`k<1D!3U;
zruX@}`WQZptAX*ZbUhc+G4)#1lRB-Q$VlY^Uaeoc()Ux&cv8fTT;I7&e(~Sh|2&3}
z2Ty-FZCxW$Xrr7g6@j)*e$Obh-DDhpN|W4>zPLyvI}}!2abB%k(s-5Lm@b(eh&d?w
z$woY-I>vV`DfQP<=4U93S<4lyU+pX@4Tsn-;V8h8=DX9DZF237$g`
zv(-bG6%T<}TxJ~xZ64WVjo%a!qY_D<hs1Gu;DbN0jSt}9>4m!G+W)W0(me3M$)frb*jsoqfOm^k
zdKkO2vZ~CTI}c2s_u;+E4yba#EwF=sLq?BbH4Vhzmj4s($pI-Z&#E^}wfJja6g-#O
zO_UCmqc!tAsao@{O}_h$tF)_0^R!QECsZDB_Qj7?!%(+Umjn+{;Xy1d=KB5a1E*GV
z)@VWmI;&@-r3HF}fr-Kbc-m$jGe;prX>R60QobA=SzhU)~9FQLWw(QR}eA~~JUV8)7K
zG`u2MONzKTH!;s3JZthKs*mkuLNT2wyu+
z?VLz=wOZG5JpWTF!d@L7R@YU&Wh3dPt3r<&`+zIH!r7jfx2tl`&$u2pp=U=C$2G)~
z2~VW=Lg@x^7S}dAATnNJYn+IrU
z*thENnp<}*uX%GOqwPtnl%5W`kKNu{myZ5It=B&UGTnwt;dszbqf^utG3mhC=$$CMthTaX@i3KYr%TnW`ATIQRAN6i0T2cUwE~lUD;Ka^H46hQb(x)b=hGZcRwcG
zBA1p`f;)C)Lt7YM6$^7qZhS=5@(Kk)Qpj}exBb&7fs6I=yUYJzBlSPaxfLGPsYTgg
z+@lY}+UYwNvWjeeR})8BTNfL6a(ilgf-S}p{2}*|&F_LFm{EQeOQIkFE|8R1+g`b5
z8uK#SZ-B2>8dvUb%L6+Z;_*QHVr)=;IK@j~jX4!jBWT
z#tTMW7c<27OO*Hsj_WHVHJmNZ!C2P%_-SZ7XpcON{b98SmfMA@7#8s_jz`#ISe!f#
zDqL-oSL_@nm4J{1xEZv9-b!h}vQOZ^^093!RWXOAfqC4U?v2_F#swXO0*XL7gTaOi
zI(b8U_@atFSa2RcYrtH9DnbW!L~6uL0$Wck=h7IcWS8PlFo_=Gyr=LgYzK?=<2U%Er{qR
zn~e`FXJ3{ZhmU&9LzxAED?-2>lanC#4{1tT?)U1alC}6H^e@m3H*!-son;bVmZty~Cm^WKu?k~xH=6R75N2sL04rdLl1eEE9$za8~dXfOYyG`MKYs>+Q
zy&vtc{hpzyz;Vp&ayDY)5cl5fXsqo9gI{ydmQQf#I@ZgD@sAPk>v*AG0DL(zVCw
zbkEpqM%e&>Ok~SGwix2hJ5i*-_Zwb`y-6=JTG!x1ftH?bj*MRS2M06c7=MB&0(TM1
zMz-j!nIF48Kv#Z7%5d3uTmHOVMfcJq?+LtXmr^4){9^F%BsMY-DEa)ydGs6DG8Hzgk``=f~=P0xl%gNo{%iTkyO#}7F`
zgG8+vHe`-L<|&wQoquD2wZ*EU@<8J}lZ|RQ5b>nha>b%omz+d{piWgw>t&oE-@4ol
zei%wbRzm%Dr=@mvxIMxc?b#&c7XPLG{9N4d%oT5e4Q63qRduTGOylNjsQyBx%}an<
zZwBl5&}c|H2fYXLI+D}FA5-4ri8b67RVS@5x{1XT@$Ze?Irs4B3CT%pYAL%1X3tbP
zaAJR{ug@xUe@T-sOMOW3b|lRG<8@D2SN~fDUB5(?yk@+?lzisschue0N8^jml-=3B
zILsORM1mnS(4wh#p~IZilB1V*?rh53blqOb)4bAe`k2K?X`E%)d#3LZJP
z_ZUfuh*GI}MA_6T_w0AGPc@2rmLSL^kiF@{wNgLSt@1~qvF&J7(@g$LXkfycQ*)=M
z1Bq`;se(_z$SGsv;vb?U^ENq9dnip~YjKhZJDI4Q6eC{
zvB=*47pzwCg1;_HV^6N$+dUEvT=p#5l(f%cn4{zLkqsLuC{o2M=E5yz11Uy)zFCy(
z@ruxun*Wn78O@rEmqtJw==Cl60V~L-@#oC!C&VmVW{NaP4>3m2d6dUyn?iYp6ZylsE^q70YDi
z63h%v5`7xZBtcR%Kcj4qgKaQ;w$cn98d3Pj(`zPsR_(5haTm_6UVM6b*5`ZN)&1Tq
z=5xJ6?ImSZICoF~P>COv>-u_PyI@zXeA!)LSRTD0)
z_H+L`n(UO$34zG&A^czC`lM!*!+Ad>pnk8-t*?WQHL>QYsY6$?^4reM36_(Hp<~fi
zJH6<6Mlp7^`i$*^<>yGSn6?zp$c8mj63`l$n+DzcxQ=ptkOR1}5Ksu8#(!%Zp}+Y~;j*BU5%B
z(o11=E!B0kEjVriRBKRDV~&4DLw4+zUt>q$l)Y4p29*xZzFghUI(P`SWjO0C@WQ#S
z1G5NQmG$D@O_=JyH!2F>NN{;J(!u56FH#1hXdRW0zipeUkR}~?>Rxo}I|*gK4B%QNum)My)U=xIcS+
zJNCaK>$qOoaI03VbMmQ(7m-G_C|$U+snWr3|Z!H-4<64{fW!g7FLSXC0c#F-{1
z$D{7}pYS`~o7)5z^9j3f5{VDMDN0mUz&^xWH=n>sJ9Ws_{kET}axaZ+0G@Nn%&_PS
zSidIHtE_Sq;bgh)+5%Hku%bap^8@T?K(D3yynubms
ze%Dz-OKr=zY}Ce~#ayLOWv9NRmIy&PNL%xk3UfdWHIA)VKh-_A*?zQ=qm*y7ZV&jM
zwTm(2n^tx(oVErrucf>LpQpv1+;c++r#rvrJinVGU7d~us0_Y%xY?T@%h+pDgGT0+
zbp{Ew#qjAsOTd#r029*xfLe4g-^$vfy8{L*$~;exkS1
zWlS??r3FJkaU`Y0A5qx)3q(|y)H7_s$0t{uD_aJd}EUVm3auJdfxbGd5gUd>~+vuqu3H(DNlr8Ur4BAzb^
zO}c+_uGy2Sexl8phJFPM!}otM#dWS395xMCW{Ia~G2LvOmK4XFtGe#S;}%?eDOKqMJmGWo6A=}qwb)^l^;X-UCPM;ZJy$Fx30wd4aj~#M!rxxN-|NsGZIP>CM5;@
zMtWs92c7XbM}pvBiC3sWE1se
zfFiw?KPXc;pFuY85WA0}hjI|ArjlJHu((9f=A%mtlr}+~ltxxJ@o8ilvY?B^gnG8k
z$DGpY0-Ok_RwOv)+}fLCFpSJ&~jbOjX|iC@pO
zR7O&m4guhkD!#7UQMh>xl@CfW8urp#-Z8`flaBW{bFChXm@
zzCkq_G=KlM5@T|5{y!=)ds@$b5^1z|?Dx5l{jT*BT9houhF=9GioLlno96i*ig;XT
z&ZD$@@<@dk`XGPl?(FQ5|x-&N1O=$)b6bTYUZZQxqwk}^#Mqq_6m>@d-f@`sx4a&
zgbj2l^JSSS_@Cj~$em!b^NNh~sF~Hid>Mgt$?5Ci=mQEYThPDF7vgw~@prQV4kYeH
zsEeug4MzZg&xYWLhwBtQIH16(f_7(B;Y0+{!K)R=pHj7A?mjk2V)qNwl&x%S!DHzY
zEoT>OQE9ls5e5)O2R$(=9U_V5-5r>EUNw&tjX8H8=UqrE={3|*?5YC`YS=u6R)At)
zg&+*VR6c2Ls+=s!Un&{aSgds7c0>Df#mxP%i7v-K-t*u`_n9B&#K2`{K63>I`Gr9U
z28GcAHen2a4`?pzqBLzKMa-x7FJkHDLFo_!6q#tmk}IY
zc=3syU*Hmzd<6)f2NELZPI-DDQHLf$;0Mm7Fs}j3BtCPLi(^IIjhF7}GY3DI7t-|v
zlvEOnCVsRPt_A=GityMv$P#vYxyYz>M`c>AB!&;Lw+KJbi7VSL0r^Q3RGrlO(o%Xd
z=zt9b-dUxBo6>a~?)35Hstfp(e-T@UArO3s_x`6yeF5Ayz(8UiW(`X0$!kE6=)T1e
z@g9c2YMN8IT^^(#@t%kQC7AM2x{=>(;X`J}VEn-U$JRLoXA*GT`VA(wGqG*kwl(3z
zwv#uuZQHh!i6^#=iS6X%`~LG^ozoXxebLp`)fZiR?Y-8skrJ^a%CZa4##qnt<0fO=
zb_hWut;fCos!wGW}fT^H5g
zE~j~#T)N0bi@l;4Q=myjpRkzTVZ8ukExfk6l6}TpDl!_b0VG*~=3nX@dSyQf`toS&
zG&V-vn|3sjpSiyJE!Op1I;ExEGPqhl|EBB3*XdjLcqUt8>Qq0!My8Z+e1(@YN}$z9
zuSejbT`g*p-6avB94S%BFzsfCNr8@Y>YkY7+^LGL;C@>6(E$@B`?;=4OW6E*GeX;^
zqVBDp`|qhdq0mGE?UAYz%#HFq$tq&DQtsLLt`MbhC$b
zTq>F!FI)>hHcK;c=(v)V8*0kb?nJcnqc)rEV-<0aHhHUACfs!8>=-I>cGN3$@PAV`
zRq{-?u}?pbSOPU3eUKrA;4{>D>TTLn=>dlYIXpu<2S~M?qJt^&PuYgsHPc5^K2}4z
zGp>{#*P8tz|7N6MPLzYFKs*sajx=kE91BI1SK?Kw2;f&ml&qQ`7q*7Qq6jj19MR<$
z>IGcjzRv{+yE4=1gI^4cp_;1^ocwyLeSkhV-f8*@pl%;tnmJ?_lDBPso*^q9XmUzmUGvY3f02FOy
z%v1@C(BRyJA*=FRLJ&Beh4eSiMM6|aBTjpM!k4AHOJir8p>R?;VpfNHQHGibVbCkh
zxxhRfSj3Xj55;!j^7OPzbmigIMFcLW6VUtHf>BF$FEN?hf-JrYoLGx0(0--tj~gS7
zj|i(t1=&iOR{~)}IfzPwXwU8N>ODp}O)33)%Edl-6JmX#1GF7rJhX@R_)9hJMErzlQ#H|g*95>25FWR5ce)&>6(*s$u-!L
z8{?%%Jl}xl??xCcG&x6@K+*jbGBIuOJgLtpy4E6f1VfBf?)aeOOYN<(a&x>LxUjyL
z;s(YIWXZ*=_%^b1vcM0Qe;c0@C+GU0tS_ng=--x#MqLne5cWj>Zff9#woL4ISF7i$
z7O4f=2+T@+K?lTBk^p)RjSG$5QxY_=#Ue7C787hj;Yh7UaB{JKWzL%+~9|^5*U}v9#-KB{KUTl<8zmI~EDk
zMh(Td+ffuXhpwU3p6);)aNW|Z*>}@-hgZnDj3hWTwv6IP%2a`0>7z{1o58iwQ=|H|
zwFZBtEgK+o6KCF!L}?4Orh4XtSlx_Sb*V=g;{wd3Z%lh|MWs=m4LaHuj=*`awA5uv
zd%L~dV89kk>Wn~7GT9L8upX7n*V$guHM`9Swt=tipt;J=HH^T+9fVpc%(h!2OYUmq
zJQBVI8+UU^onb9vZQc%go7K@~zUT5RhQoy}ADlD;A1aSl@uZrXuc9CdS}htK$LouF
zjH)hOHx7W8Po4Vq|Kv@HzU=l2zJ
z(VOJimPHbA08(Jn2F#p_aqa6j252c61(LFEAmYF!bZfT$>E8Z4JCz2!%wFBXE)CLP
zLfDmal^2^l!pk!|e1x!V8B?Zz4IB8;8m>;lb6gw5i=L<_
zf7^&n@gf9yiB{+NntqcuAK0v8Kajw!$DrJuXpOW}n^f8!VvV28d%zw>uB)8@uDZ5L
z7fvSS?F0nnQyYG@82_h^TK(zC-dJt?T(;pq-aX`NpL`oQyG&m#+ZZgXyR@fk-zx68
z5;EHFGHW)bRbCrBA*6^v6Oct}wN}~*`yHI<#%c?`(?NEK`n-^yRQL`}yfZBJ9)B<_
z)Yq8tc(18n>^M}q_*VWN{#S{N=s3mWRNJo6#SP@CRR8>l_j5Z7V5;{yXs=ss=0yw2
zN5E)VdZWKRByMZegj<(Ib81hep@H(_sjqYts1ZNM{Q;Nxotr~-I8WqM90|1@xF^?+
zSbba_?~+Q0xbqT}FwLT3IN=+4NPM^I1rzI0JqC+&48tA4o43Qp_1E;>H0#fywa>8k
z;Q^kc8xWV2wxSkQpKZ;^*xcC-JkmpW;UscJD-3@efVJSQK=QuCDE7kn{Y>>Krxe!5i
zfLwB-Cg~TMhb{JLx2z)d;EC>|n_uTD7pQbUVQUfrKhHhEZfn0FZ4QSl`=R^nzQ1e|$T||1ge9%NCkSS#
z>!fF;wH^u`S-8|bbN%1#3vz;gxl%At5$w+%*C0x>6B7&nosOi%Q5YW+7DcDepQ3Epz8OQko4Vow8(YEorwaRl_9eL>vVDZHOfupOfx5A5XfK%$YpN0VD)ZRK$OwyQB*Q?p5UxVWoe#D~4bKW6Sc%F8cr<=VX~_1xn}Av@99akYvas*>mfg(_g()DF2Qo0?#0l_j;ubsxDSxrgc^
ztrZ9e0j3Y>y1!K!a<9T$Q-F}%Ah-2DI2juz1#f`ZBKXo_2GRsK>YmfluHo6!7B{CpvDbOmMJL59c{=Is?jZy-p)-c
z*H#;JhMCzCP64D=d_T^?rn^46x9iB#{f*tMN{
z#|fXWQ$1MLjNrX-R>dl4a5%g?ca!InEc+`7r7iyHB7$SYpo*8ZS>J82jg*Xxu36|ui02nQ1n
z){gNQEE@Iw+G3=(QVH=nf>@6~Kq)48hqFGNOex<10i3jXVVek?SV|udU@|{+bMb!I
zg@KxRv0mEp^cfh#6VeXnRFX}N6)?ehNyp4U%>kolnfAb8ndKy>9N;i-f^VkhTO$qBe((TImD^%n}^5oAq{;4iCX>eTG
zWv!PKoSa-8xOGKYh0%H{Rl)E?W5;3LsvCdby0X*W^hcd)2TC5RVvHmOy?D&_%FP>J
zI0#5I(p8p8P4imPimh9No!z^s_OMx!MNXQNax3DMybzvhipfvs+Y7OD$x*eAU#L>I
z#WPzQ!J-q!=Ig1Gm_dzImUU`@)Nt69E9>Z1X>fJKGYT~0!9vz4WmWf$&`QbP#uK9z
zRpcf{&=D>d`|ouQG-bDT56Wx
zehJ>QDsj(+Ll8L=mJc2-P3c}qk}b6hT!I{f#gU}hTr+BX29J;*?{;;Y6V6U`bpKg-8(u`K?Lt$l9Td+oNUSZ#VgSfCw*7zwFANpYpvPHU8PuVUTlncnB|m*_saQi{AKCj>E6@-BpFV?Ojv0g*3r
zE}gJAi+n_asU{TEPP3>PWsx)dSoc$$QNg@nL9;x_98M*50t~!$=yH7#s}ShQl{RHZ
z6Y?uip`kV_xdR8QC&3>T>b^i@@&2QvW>d)!H|Y?KhAkwjPnd&J^%dVEau{#HVa{ch
zScnRzx0c6?n`DziicG$jZVI3YaXXwX(peNVr2pNKC!?kjhwiN!nh~1fEOV9Ew#igU
zL5pQ#Z|m(ElfUKI!?#HJeFr>=NJN_nJ<^Pox_5>3(RJirzG5Ebnyq+vG&FhR4FpvU
z%Vx55)deZl-e15y4m+%0r&d;ZSloOdOpac3Opgw=s4~wqiMa
z1|RYs98YKV#%a~F%rV
zl$fz};hxSZ6S*wY+98k@hvaHp&o(_q9sbb^E1`S01&Ia5rV4I(i7;#a>zx8c7X$G*
z#}kAkT8!gg;xy@<>H`oV$%<2qyGrANrjdmEQyVeosiw4W9}xl2)0$gK=Z<#CU_
zT>_)SquEc?oJ25@VDXz~0r(*WF%B^GoR9S5Lz3W-Xwy`R7|={?jghC)0GJQO*3}u2
zS5KDy*V*THvF5?81>cyPsG;Zw-@DlxZv#3zJ>=a|goChiNIe3LAXOYy%`kI*p+eU6VN2-%_G~O)b}EYWgXBRJ&q@8uO_Ex#gsahAR#~O
z_Z%OC@A#+B#VS3~tGBu%o?nhW7dh;L!!*P+O|=?2wrfRyhNdG`(ozYhCs)2%KFHMB
z{rme$pKOuGU;X9t)vBSCCJL;)Q_!!)NFvzo*8bRYeNSh<#!Pp&ynvm89{7XZ&uFVq
zT|XE~8SU4e|DyXv`U??jpGsVtuxc>@eAiP5`=kIzPB@%+wu^V~sIx~*H2>qwY
zv#z&d)AoGva!jZ7Y%&tgseFhs%AWmCZWxEIqWQd!z(1?z?v3smSt#K8OBNo~3z_2q
z(030OAoc_QMG9PCX(S$uPYH
zW&Syva*#M&!;Kj+F%3BXEP(LW7B6zAt|!t7swLY7e@P3M^4sNU{#L)9y}d*8TW9u}
zk8fD$E0~}1Rt*|-Y%)2T5v5P336y@1bF$GWxx^z)?D~|%`!#y4+638~_u;6>@~fW{
zMpY-Xr#_UVYn<7ANUh|HbLJdFKi9m1T}}hlV;e`$4lx75vbbL(@#1o
zmfFCq_Ot836iT3}f`VQ`|CE2JX%5J}#7@9Tca54pymu2(0lCv%LgK$gB^@(Z%?0P2
zGfFK-I$
zlRux^*E6?b@UB+_tVa7@^qdVjk?5p|Odr(?^@|PEZ7djZ)z%Bpo(nquVI;REMiLo1
zGT9Y668%yxGK_DuNBZGX3(4*eMZHhEIGef5p`#35oen3phECrFgHY=W*j7m3bpJa-
z<=KZhykrY}Rs^neru&CCC^i#{U>cjJ1`Rr5Q%d2778?h*t~uP*g~08KMi77MiB39&
zI8F|340(@{f*{Wuc3Bb73VJ%<$@yGE3|0g1&!JM>3pB#$Zv`n>$>BgksuYjPNcXu+
zJr|vmAa|?h5#KEsezkm8z!`Dq{P51Sc+!f$^&Eko^s9F_b9U
zl<55*`Pk+I0&XYjU1>H!+m1EMaFFs_f95>m!S$MK9^zwtw>=!jO$^AWb|k}4D5eZz
zE@=QyXqlA5td74{j?
z4PKyzRze$8A?p)pAHL+7#NrD#q#fp>6=4Rf1up9eX5c?(o@#NC0mh>`M_X}ZkR~kv
zoG6jJPq$-|7BqzCqAUV)P-_7>a;2zge%K1a0SvPOsAox5m_?O_M!c#}GL_h}baF
z`R9%UEI#rdRhQfHe+HK04Ybc&_Zb)q9rN9h29#@Q>BVr=)e4-M@5_Z@+))GtoUZ9u
zOIPt4L2S>=u7PTKv6+ix;l`BG>Yu^X9I6HtAqY9e0|!|ZT{`^(g}1(;uub3mT^eAG
zYOI`C48%P8H3ttzZUr_ZigwpeTf^BCl;9o$D5-?6{v93_j=&#*(@N#u_{qeVALZq-
zCFQv9E}|@ttbibzz(5Pv>D`5PFo1$XMzF)M=9P;T$n6p&M0D0ZBacGx3YZEw((8`V
zLp~E63v&G5V>V1-AKv`eSJM$}0&}2gL?v3bHkjQ6A+4iL+H?
zs0qtqC%=pUT!LAqT$@OR(~w(EHONEKp1e<}cF~~2{|R@6#=gWcgl{(T$>Qhw1DSyY
zb#g~eY_NN=L2&&V^`tI3qY@uxq#ybK>7ll4g`~1Z5&*GvSPhAAn;H2DQWpxR)%{E9dtA%J7VY?RxWb4ON%)R$tntBhT
zaJMF0fD(k-H^#ZN1BUyq8c!HvA{ZZ@pV-i|6N#Ioiy@*y+yG@#F>G9JcKiwvtUeBjP2G?}|2-lJv5o{2j1v6c3O)s4
z0YH=>tE?XIzM-Z2;W=D%0m4r!UB5!Osd)*pL_22<4)K`)zRqYuBp}SZ*T-|#_Bb&#
z3*uU+gF)a0B}r$B*>X%W&DMh$Y6#vEa{3B;Hg~JWLa1D2AHin$5vy=_Y58LoTd~vc
z$xv1?DponJ;8zR>=mQh_LE7VSx7DrqoTKO+xL>lH#O{GMs_XCk``@0^
zZ3cItZJYte@j4Ud(!e&}YnSc(Nwn-fihOv5bd9Q%L55F$V^Q_%3P
z)lWlBT_uFgB>e3+m4X22?09*UHr~~93aH8b;!Mft>>yEHP{6x?u7moDyQj*2oUpKn
zJ!YQ)K49UA3ZWPr_wdye{@j0uJN}`*inq%k7RAlXCuG^VhfHWKx1}z$VWwgoYmLN9
zX}Uwnk1B)HLyWK-1#t#voYVlWf7~A%khL~Qb{t@p{!$B0AHL1Mj2UZgTB-gS(i_k~
zv~<{W3yT?=YX{_&?zxE*^F)SsDpx+4EsirhMdrFr(Y3rQoDb)h8(uXW%k^Y#K39__Z9xn<`BbNB>Qxza|#3$B+i!V#p42
zTqX67xT8E22QNK6WhQ!Bkt`mMb{5|YEP>YNr|)SF{R_--{xh27^FcSptSPG0_W3}q
zSp*rbAtl;@g
zpfkNr%8gZ+4|1BjJH;&59pm9B_oq7(ZK1J@|>P*5&;L!wvv`AA}bI*QdqIHKrSb6O~
z=L=FdyiFEp)iCBH{9+t&t>aFNtqYKjYhl7NZXG#As8j^co-{#g({S9`K?2#(gC6_r
z8?5vQXTfCy)@p%OJr)SRWnHU{=sNo^@U~};h^HIUi9s}4^`sf-5>ky{zp6gYJ}SQe
zmv#>tsr}7hs3460iT$I3voN!8{I|qS*4Bz$A4T?ks-4N?>Vw<-YpW+8Kq!7597qVl
z-nbGnxU6i007UjY*IZ9dJ)!jXMlW%1v@Cxawlj9yI7
z?o#6V7Ma++_vKn6Pg}mF$zJ;&^r{lu(+Xvw<&J-0VmgBwM^sIXe*1dUyVuD^O@Pi0
z+$z)e=x}`dW-fwYffHbBv}VQVD)SM@v0f_(FVn=RWUP2+Bxh|wG1(V1)F7ru9#%lXv7{a)#~#T?X9`T
z90^g=qz(_yXqACs1UCtrqbVDw>a9q1%4(jwMKTtf@+Ph*Xob89CXeu
zDhy)(7G5}XQut^FH>x%L!;I>vrng0uxF-+N)b2E~NFx@`+rEQXX;&p_*3#$AB!7sn
z2}xnCNqq`&)si{3T=LjPVQ29JET|k@WwaIkInU|B0*3B7AgnX@L7QQ9*x(7FZA^I|
zwiY8NHCGZHC<$}UAY!-~Cvw)n3z@wmsvc
zUeigO9}>)L8di^!);(t*DWonBr>z;sBw~ko+NXbx8T2Z|on8q8$;d?H8K%v#OmE>#*^@Unr%2N=vqf75T>@>m_
z_4Qt>qrYs##EZTFL%eP@Dw9b2$~bOJ)=S6@BP|$emNAETAF3(n
zDXt}rjfFitSq?m;u7#qM7N1{)KGHlZ05xiKPH4lDJcbNxd3Sn4IJ`k*3}Y1`Z{sq3
zaI6cOK%yDXiXt?eWwLMy(!93MO>-Bw0ybDPAen5dL0GQaFGN
zOyd3|zRq8P;Ri@KPExs38=ze($;Zpu6QbE;4*`ek*Nsd
zzhq3ec7!!cVdRZVnu2e=rPlu(+P%^xCF5LX1ca;D
zbLxJ|yIx4O$V+jx>Bg-|%>hUEkzHii`o?;QHoCn`DCms^6CCz~GiCpHP~UDvv4QFI^HpeX5H#MNL8^?U=)mG-F2|~8Z
z&p<}LaO%Ft4hHP3QY7lH+pAVhq&}D21HDUz+b@=)YwPwZ#RhLl
z{;<=kmL8<9tS{VAnbNkU=8?E5=iK;A@`Fe`E-2h|DLKp=
zg{6=)(J3cc0vD8VEEB_wIP2XCH;y5|(;jf<%nrg>-fUi`pO=y#5#;o(Hu?vNuC!PdgZ|-gyI9+2Q*k
zbN*7dbUJ)Ba_J@pzE}b5IIsc;MTw)MH-F)|iHf|s!x^WIN;w&P9sj0tZx0>qFmGa0D*mp@#C8Tul$a{fkgaRpcZ$h_$f+znn1(gRvEeUul=GwvtT
zB4Y;WoM2Bc!t_xs+8!74XVdj4D`uyT2nUN~>AS&(k-~_yF65Ga4}qs5%J38wq_k!U
zm?n#7bN}s9K2X}U0Y(L-)}eMdJkcapz+X`W46~D7_3Bw=p+qea8C|ZUbLmZtYsu96
zu9%ZyyQ&I-IrI(-^2IYD6Q%t=YtSnyV(KV(=k{>{!@t6fhXNILb4oRP$xbm_M`KE<
zkP~T9X~yOk-SxYAUU_X!1zSrq>!gpdh9D+Z@Bqh
z4STj>SOhQ$uQ1%*-i~s$Z8aGJC_E|{S{>vZKQu8itD+MfT~falZ%fqQh|Q<%b`m)w
zo4ZB4SDp=k;^Zwl$8EX$&5uAl`aw~-;nrs$)~#Fx+}_#4k8Psyonu)WJEVti?o!jJ
zr_gESoB&XMOyrOrO=nr%*f}exLl$@x8l8Lp{0GH$`vyrP_5jBM8Aey7fL{C;eKZz3
z*UM}84Tm?*rmM}iz%^uc$C}*XYCya*3SohJEi-=4AV?o@v&SNnCG3?s?NV87*%x2L*>_<|u8L;y!diNUy-%
z!(}8Ut^Zat-cfun52u)t(j@H-p})l4kx#_kfFEq+tGMiAN3~ZdFCTaNort+jYn14z
zS|uE@Z_(28D@T9Gr>(2zt?x8#KGA)=S4!9;hcCOTJ3>{(wsr8i24f?Fhw?bY^|N;cLbH>f$x_ILP=3UU@oT7&X15bg18ULpO$Up0_3E#_y4W
z7qXaXvu@ek8zDZ;A#0`R_&eaoJ#V>v!h+Dgu2rGzuM#8P4*l%wW}v(6M7D4wnxaAT
z?`{$3*g1FzSeudCJE(L3KY#nNf%{3em;NG8S0ac!aep;bjD#mEL#vYW8s#YYz>eEI
z>U4p%7Z<`~Td75X;dS+95)5~RaWR~=fMlSvpVZRYm14YE4!7&M_JZNEZALqX-Jxf@
zB5zgZ^hJro`bBFX4=@`imRx;@?rrwT(H7}8MqCT}vz%o%%oeeT*T2c$@et>-aW%oD
zU)!zaYLEBJo3cT2Np6t%{v4U{bGY1a@`>+-jdhc-_uc7|rz?e^c4ITtM4(>Pbb_VHI`vgVRDU2AF$aIR-VMj(I9k2|h@wK6bq5Q?u08j-UG>5w5p_;j#@=mD`OcFU|W>a@`fHK7?&tGB8)b%z2MzUH5ip*X?%&N
zQ=T@cpBEFDzv8fZg{rdteT8AEz~YCN+u;6qP#L%?U4sXyQ$c7qAnMovL!=ciihJuF
z?}h;%v|p7uR-2>ul4Ym}Kg$RsjGmt@u;hib@;v-o0svy<$+TRXIU6*N6NX$5=6_Z;}T8v^-B@OVO(bBC58!M
z^WFToz&)m$BRs&71qd9`R!T){*w!}+{a$8+j{_R(JGlMSWHCpO^Wog{_)S8fR^Xpo
zjsTbQ?vAz$x(s`-q@PLlcaN7!6IQ(5nXXgxH(lLyUpB<$ZU9b|iqKEv-YGPYo=$aF
z6JK@;T+{*Yv;!0E-VsOv&2Y|b|t@a{!Hr6xDTXa%s=K?*>b0Ge@D$=XoI0r_-I`;UC1^mA#e@k75p
zQ&=--bmtnPs*!1kPI`L7)|e9n`KGa&7JxFXf@Nrow3svcyRz+h28W?eeB{Ad>N*>zQ2oJ1WFUo!I9S<}l;(bGgbx?)Cb5gRpKRXx^LXX7
zuPf?&pnGn5j}L^IvY}unJlB+pit02}>flYCRDpRzY?6XZ_?YE-{@{6tm{BuO;=n+1
z33gUjLUdIX5=AwCMw|;SEn<5-lhuf~{TStSE#rLZikhdc_TBEv>G`(0N8(LI_=pz~
zYCY(2*NsbLTv}Uwy|!+I&WcRivS8lI!)w711>P;Xb+{)!N)^GjgCmA&344$1
zLw{|h5H#hFH_}Ch3KK?LVW3O?CU?^yT05d9EFcSOs7c)yo(gbbMP1MS{
z)beap1d9e!w8%Sd>3NGQ3gz&gVNsOQ^{T|8AaoZd_lHYxP2jaG|KtD=BDnw9|ABzq9Z5+_Jr81+s!F9z9&oCx8CH@fN;PVi{!b&#M-k0Q?joA~11i5*!;Mu)1X{C8jo}IFr_6lff|=hZX&QZ$)K{
zWMPT0>xL1Dxl%fZTnw%sLx3j~Ni}VQX-YdhNx+oNu≪p4|BNo-K5N&>&;X0%_=m
zI+ol9?`Co-J7~%uJ|G@@qt10(@XRxA3ef`bJ>xZ15&90dbZLQwo^qU`xOmI=4Omu%
zUUOzU)7ta|vkR0kSuMW#^RNDG0ZOWYF7Rp;SSL&3!0p(cVrnt~IB-cwd72W!7t6ya
zTf!HMISTlQ@)gB<-&d8`Fakf#>(|~tR&?%y_{o}dCPv?4KRIqm3nLxVH#C4COpuQ!
z>1cGS`n)=QIvX(2vbfY5+VKFmR@2-OFB65lG`x_t6xNBcV0c6|rFmIn7VNV_^pauj
z5a)yMEk+4IjTzkI1^P~09V81|1ySz8_r5ApGBEHW^|=UIQ%AHZZ{212b+!lAo8oz=
z>enEn&87&k>Nv>Wh5X)H&>~Z`3Y$`}h1fr|v3fpr0_y?A0o#Fl%8TfnhL6L#;^pvu
za5e_%=%6r$jW0ElhX}WLM+kF-Z2PsbVBYdZqDg~Z*uDt<_=?&^!~io&Tfal8xLjM5psac#gf?jdT=F_1CI|A0us@dDmvc^?^d=CvvlcU|~>nFWIF
z@qF=5h@tQFJ7gSVR>b9Qu~=99)(tV+=kMZ`X|^Mg1_D*Q1nF-g{zUO!@!sk{RwYZr
zKcVaSopO-iN7OqW6u_K#A_$jP$CIRc-v0N?XqaYoH8l_E4Wp@q1E*ByV|F
zKV{-XC_4VjR4nXDR%PeTSaBO|$Kohh4K$<_xV;&kUp^FEr%;BiWrdI)!!IjW(ZqF5
zuyqVFBrK&UqlC!jUf!NJhO3Y=zW>&!SWN%FC@U++|F}%D)n#H2SpVatADyL>Gu!73
zPb7iDzzFYVM1c$mq_}ae`de6^osH+qr$*mNu?dG_-=JI5aCWm;I4Z9d|8ZRtcGmiI
zwFTT^KD(`<$~^z1qPeXJt|Z>au3hhncUf9sW)REXoL;rF_Tk1AQ>(k;_i=1uF+7w7
zKB!N7dIE`VjZOC-kUR<&^13=wekMor`WCA_0Ii2DxvSXqVltj@ev~3+y*&GYE^S`{
zJymWhrY2JET^lpV<0xq!6&L=vQ}VRX1uh1XSU%&cFVv!)O9@E4w%N+lfn{f9OryTn
z;B^h$4c1p^KRQTNk@GJF#<;a5D>=E}cykWqt^xV0UBACTEa{)lf9g&|j!Kuh^rSkw
zRL-I1D^)^t?zJXDKK|R33Y~v5$KK1s9l-y_dUBY6b;3}L59CB0^+Sg*Noz!-d
z*O)#LQ~E{R>X6Cgz)Eow<2oL62%ddF_Pe!l@2MM0Akh3|_nQ&vKa7^ux-^z7B0M5k
z_5kvoiixlaR<^)MLg{Z2rXBylR9bA<^`OeEqH@<$o6X&pz55qn0cZihIP|lKn$J+r
zI>0Adf0dD4mjOO&vuD;e`BRF3iWS_FVi1UW3CFo@b*>PZ
z=tV}H10KlA5$&PrtwhAraifKXI$f4Tua@=};4$&Fyxyokm4&E7mS85%ixV*|idWsLpx4_=t
zrs}0-A|pL%LbVRHC?jZH-fMnWF?FDXduUqTePk={-K5PE*-HT&TDp;4RE2}NB7`*Z
zCuQyvRS4*g02*VWKy*ojEeN3I;X5978D^;&9Ef|&R@(2?zFU?*Pvm#l9>k$PRF={o
z6)T(WVh|J~x2eHg{@j&>)AMVwVFBelamNDhiHx)4PXeVa*mhL`ZHOkIyt?!>n{Y?@
zNOHY$Nlf8B*yN2_C*U(nH)Cn>f?d*eUy8{YUrNcoA?A_HloUL!Ei{~-uFlH&;bRzZ
zxAC`O{bpxJsJ-~Iv1wkFaTnYfarR}ERUrv9G(tTHC&9wq22`aPi;T;GZ@@Rrk6Z>l
z8+soU)c2On1+mnL7P$+EcDXv!7N}rH3mNR{~GKNG}wzYU;pA@QP4P7A^_Q;wq}|)Ww&DyA`1jw&D>d
z0*0dl4F9ap1zrDX+p#o&r6;H8AG1JgzDKJ_HSUdpX#B~r7pHWNJcj@9b}q{K2F9dG^^)Dya}-dwl8j;)a)^6^x8WdJO}pKf6pt
z9-ka5{X3LG^)R}a$Js!A;Y!*wcG(1zO|c;el&JG@Usq=+ccGBRhAqHRaxp9%M-3<0
zrlI&UYxLGQ7nT}QhopCu8V}=6S}=~=kJ}j}+eAJs#TJiQF;rKJnbC=R8pGFM&ZXuc
z&XD1V)JhIUMnH!!mmtau3VUW`CZjS39apOJfCnrB)zc12FNoZ@?G(h=@$9`Ongn;o
z;ZFi7kFA4WLpfnnMR||IkLz8EBP>4OKI0qLyW2)iY4@zLWiE5V49D5W`KN~oATj%%TwSQUDzsU!Vc=>%G^l#FwNKEmv
z98Xp~UA1b?{3`w?sq4CN#Ca!c6d*IUuz^Rct#QsW&dfQ17MY;gRKr_iomha|R7ZhC
zHQnD&+LMwBBM5>~IO*URf>97s`28J&BQ!g>-#q~s6qOzW{YXH)x;&9%79N6Gxo5R853M>fXs5s#e{t1nYtIqBWY%for&&85uWd#y{++Y;_
zko<|IwbsF{odE`C9R-rm#M}hX(pg5#=knWFf-rz$WvOEZS;+t-`Tql;lOc^l4t2h<&XO@`(#A
zxV8ahz7O@jPipzbQ2Zc4Bap>^Apb@5I15!Oz*PNdf?y90`b86+?!T|!Gk1GC2e1Nj
z!lB5E_K%(o%}?}CY$Orqm7ri61ydwjf`qt;CmJke#3=X00**QRbyZSu1f3^WY?WTd$pE&(Y
zbs$lHqqF!!-uDaoKG1ashKDfvZoW5cB?2oyub#)gI*QN;I2KlxhR5*?;8?%M{J>_H
zS1gFzFW}|M*8z%4OG5%%9Yn&Ho5=ZGl$b13!8*Wl`g(!vJ3Ij!BL
zvGHSV&8H0guJJX~Jt%vLdS5K+i%ik&F}{`$-1*2H(^q37fE8}C>Z|ZAO^ieaC$Esi
z??K~4C
zj)CDvF*rhr1Gf4369s@*9bG*i3fnieI)-ClbOv}jCe+dbID-O3zl6^L+`C_d?4E$^
zlkG5ML5fcfBZwKP&uC2$=Q$t=J0N><`-G(A9EkKLf%c~t8P!kz)b1V`sPbZW-!FS8
z5VlWJ(*K`3cDKX81kLUVhHmp&8k%2EUVF(VCxN6(OTagl8DNs>JK`uH59||?*{AN+
zJU!X}L4cOuJ@XwNTR=G@?Mvii*XR}TIezy1>YX4tt$jTm#OBlRhUl&IO9G7U6GWh*
z`5p3|a_k?$HOY%W857WZ;Wl;@j-Ywv6U-N>^(}K;0-q-@u>C?H!t8m+_6za3e{+}M
zIixfN7&xoLp(07PG)};t<4_k>y_C{0p!aoD^6)3v`EtS?lnaV{B{$
zHOJ!81i)1hm|f5QB)|H)WqVy_Ij!Y;b};;k{^CUf25lh<|_;|@8#s;-+^?-KMv`Fp>7%K2g!y|m2jWpVRMeX)ij=eZ3FZ6j)3|evGMibdi4I&xK}e2VUjW@3k2+c
z1g5a=Swvprl}9X_*{3HVrOnAT7FTJu?~N~I3=i9GNds9}z?X09Aoh
z=zR984NNhMI?v1JyrC6oXn>Oe4lP#cNt|*h$fK7#3bvJVZ(SnVAe2lUSN!N
z-Ic=XZ;^iRE0v=cNLAWzPMOVtC#n3wVxW&pI92;P*L~M`3fX-
z;I_B6gX`$?83ECNYh#^@*tbEpTn4J`S(>fz1={W=qTZHh(}&u;rsXn!LBK|-9oRd<
z1bO_@{0x18IsSSeiekjIb^nGn
zS75G_7zh{W&QDqu*h*FOe;MQZ`M|sbR|YDc+QVFQLBb{;+@BU;FhH&Ohq#JM+fh^p
zw`O<@Ns~-&$o~UEK)t^ik=i8;dIXn0GwJiy{&X3%8-f?rRJy*N_eXu*PH)bK{*DkE
zq^&@lO-;z(aQa!79>JIW8`SQANMh{n{YXna!zqK}7b7in4EC$|S$yK7nb2F#lrUI-
zp)3y!i=i7Ya6;AAv}CuGJ=Qs&Jd@`(>NvnSHq$CFdmOCeN=&$QLyl=x9H*XA>Hdm=}J%eOYP$
zRsb!f&Is0F(Ng8sILA~UXf`XWjy;cm^;;m>a9L9gcfJlME;hpx+=bv88lLnPu7v0m
z^f>Kqte1x1=Qc7LGQeA0^;P=zl|-m<-|02z7x&gHAyRcLGvxX4<&o~SD{pcElt(gG
zp~`v_7guShR09+%xmX7!Rye*+GY}MRQ<$&4ciCva9Tk&a2H8+iNOue3&RgJrH^tOD
z-vxLISahMXppZ4=9P&%g_thhSHXDi_X#0EpiPbf_@F$tc>%=V_jJ)jW_QhXiAbdCS
z4F!vhCkY8e4g4mex*(-;$(<$C;gMl;vK2Y`S`DYMNP8-|y^>@H4mY{Mam5TPrA)hG
zRovL$oDXk?3`S$O7|&WMgeYr&m3T(1l|+jNeoiCa
z_`crYus3PE_n4VWK@phxsHRfk^QDg{ts}cK>QHE?2PvN)fOORcf&P+udC9`ANgz=$
zybK0^PmWhMqnsXoXe|Ck&ZRZ+GdAPSYiE&n&0HW|)@N~m;jx_1FO*Mz;@qV^2R+EW
z->&FgvYZ!iz}`ohS6nvcRdWtpGVnjI8(5T>bZVI@pL-pBI_zyY>a!hc%V#bil{E%$
zJqTolPf{fPw$|PlsGD+P5+5US>@SKbEI{g`j1O5;^>h43og#l`q>HYcbl0`AjF0OO
z2ot&>^qQhAadmK~bIk95pABxCSQZj^X;kzEi}=*V+&ID3?Mb_M5S^&y2HBZEd1p%a
za~I`A2#&|;^Bo4hxC3Z7n<;83PhrKTPn)P>pCPWC3}1|>CV#K;&pZuZ4_>PZc`
z=Z~d8qsYvQp~i4aN)nUTV=Kt~<5EZs
zEL(wOw>%zWBxllUa6@5(5}D5n1AMZv&haKQ+TNzeln`9evy<7;N#!Bf2{tGBJ>tuDa>qJ5QWWMR#URC2{M&4dlYI;{58+m35USRI!2Ry7Y|6&5sk
zoiCx7$}pM4DU1`BP;JhYNVQ|YKP}w2_3?BZ6pA1O7O{$DUE4$qq2i<1Q)ktzL*fQO
z*+_>6Vg-qR!*QCf5!%!=8VaBD7kSe;P{chQ;Vh|2?pT856^ID
zEHyiXcDtha&#v56i(jTr{K`?YRN{rj7YB6H?fNh%yL|_HW94Ru<0aPJQ`T9m8wlrf
z7wu6uy(t(ARi>2jqU6|Bh&}-08WtW>;h1>_au&3IoM>8sZmtfVuDJ6l-H!PEoY%~H
zzz1asIbGtC7>Jn;SGNSAd-Wt_@AjYA`+i+$*gfnT#kDL!{>I
z*pXv@?bz{iT!Cz_Qw{WKZfO?_s^+q8h0
zF|58k4fV&0ikWp-;M1BI{(^XTVbGMIP^Tz}!ivVc{wd>FgWi!u0A?~AY+i_^HSa30
zxh#zSYbZ09>Wc>=NP;wJN^d6wZ175?Lb?K)53O#Etx`P8zb(!pgg)0*!BF(+sTVvx>oLHv`LffzD
zq~;ih+Ec{m9l9M(4q9{KaFLo^a`9e&rPl$2aCHHoYQ@W3mOzwJxvPAZizx8??HwFU
zG2p>`Ib`1M*o>(+3cQ6O#K`y#SE_u}4
zKv?x4=Z^30%#+r@+t5*!mYgkrm#0YhedO9eDMmjsq5M*~H)EzizfrONDB8q~?3r+rgsyqsWwCigN(S8{$l+NaUUhnu~qowf+P(+b?&ICoJZh4ERF(D6GXb$8
zs+g3MkO`V=S5fbjzXt8*%?Dqq^0&F$Eq1T4$5(M+(xZ47rH=>R_Iy$E*+
zn06_5_{iLiwf~j$DJAvjH&EiOZi0NpHx^rMHV7DXQcywf9rhNXQJL)8`q=KsIaiB6J-#s_6&xE@XV<*`QUvc^<|wY<%Yn6
zZ?Z9VYkhUs_c*?DRZ~4i_DW%yC6VKX{veBFUWm3m5z5R9@3{EqJi0!
z0wtO~YrppSn3^7ie`FFCIX!I!wYyuv*VT?_!-qS>LMB>$*QohBhp4z
zx3#8F*Tna+yfm@VrvWPzgrF)ggJW^1iTtrY4ohm&wKqH9$+=vTP7wnnT0m)W<*ms6)Na7zfOA3*=kXYUY`i?1PgP
zB!z`_5%aQP+oGcn)Mq#Nvdk&$8u#+=#klVUKb%Z*@D+{hiqD^E>BdPnG6)n#d0J!d
z6$vtPMBMZ?)e^N^{F6`%68yJlJv{B=Tt5;+b*kHc9H02jYA8h{Z6_mWfi|0#k{*S9
z`-lSt+!KN{#4gPBFE7bb(qxuG^L9nT!F)@{cIvTVaQ$s`a8QwU%dgklBnt#pC_|m;
z?zfvrf%A;R{n0++-{cg2Gsuc>l(!I+Zgoy!UVX28?jBvPhHky(-ng5||E#?+T|Dn(
zACXaiV@Z!y;Ep!Fn2a#Hw*T6tmcCqxyVBz>MPD1k5dtIdx{U9XIeB)!820FDZ>2$9=9a
z?+`*$zsqt)g!!+yt2MWai$z%AhuU7p%Sz0rfECv6Bo4+t%Qd84Z$P>;HQ+WIh&MTZ
zFiuMenn>KRv$KSZq9~53ACU*DzEIcky-4i4crI>E<{<2UJFWW>;=pG!hdx>hNtI-T
zzcLfe#w|VaQsu{PZXYpEwio~#4eDe{bW(LG%mu@HYHgj?v^P^)Gv)
zIYIql*hUCQ_WA?DVtRZ1Z&WkFciLNj$|1;)Qg_Vgb6?RTvJ!%lFa9Q;vh}Qvspl~#>
zX~S_zg~pa>taokFn;T1Dwd1)`d(Cp=Qwx15LRT;<=(S0ptW85-sL
zffa*iTceY>eKkT`*d!r{bOw_MMq(ZMA**Y9VzS$T>42nTB
zsya^0@7&J02({hyJ|iU$>m=*m-rSW^&KB`bG8F-QYe_cS{Bma+XM*0ZJg52k=;c=MaNjR+M0STyTuskczq#?7IMw*
z2?A4H`m$`)JT!)grq+>JfA`XNO5}4(_c+7Tcm5
z9ge3@see<3qBZ!FA!*j_^BNYUHZhtC&@I~vM&>Zy$2PPUUfsC_T
zd=Epq+YJ?5eY{^T)M3g0PW)$w5?`Z@&)B3FB$are|KJdz!|_4X@2;~o``Q?H}2hI>F&OB1w+c*)xJKmtLj44sH;}JxRB1#
zdIHNl)ioP#b!3r3x?Ic
zEnDjwc1X#%5X7|l%s?9n80CY3hbc_NToG!6#Qq$AHV=1-gRY24RncTp?X~YtK|cg3
zt@mc?y0GI^htyNP3lrME<`-e$9wpauA=beL(xS?KxO1x>O@~V|%gz=c{IZ{8Z##z}
z*@Jidys94JP*KCX@7^^!aqS8D=a|(8!um*Zpq9#g~hH;nKLb6*r-n4i)H@ZohZ5Wu>oY@yfkr*=O&W
zlj`EZLc1KF_5_#&K2NFaKSit*09-RBSb!$kj;1;p`l^q(
zufg)Gd5&c?0}Ltg+hN;Tn6>W^_splmzWIXhU}>Ibg4w!J3#oOsO>7W<
zr;{^M-FZmMCT}w2G^m@4!
z`4X1BIc@=1x>)1T0Z4Bdzg8g`m
z_cjuA;0EZ1uRr*b`V&P{ImYLc)%u%hwWIx+0k%3=G#`-;djNvD{Lk7=dt5AkUOMAY
z|4^`H7eScSG}KVq#nIQCK34=2HysmPoB>!w4U0Pr!!!xyIDPioc0Y^64x49kNA8`>
zo7kn&4asViN!63Fl{_$rQFo19HV*HUJT5wDK@?tPyoq&`hu#bqt!BX^_mZ3q=H!)<
z7)H+-p$pY~*V@gtp-NBN1|LIzFp%1Q&1xY`0RpBug}FniP=p~J(C@Pma29UssXIf_
zGWob>Ptz;~(|8TVyB379b!laID)rEX)%53zwJJF(Na15rThdFP{lh+GQ{sld;aOZk
ztR=0t6BG=<=b&9v?$D~ynFt;y39ZN2@-q)Ks0AwY?3y*YQHhpL$>i
zKtW(DhF7l{sYu(C2^p$?a1vePbW)((x_)=lztAfdryBSPKJL}wkaAYFz=>(OBGY|{
zvOzI)^Hr8m7+HAqnN6m7gK&M*4t*Sxa?=H5dYRWGc7u-(;=t6lo$_=JMAjU1nLrSV
zy}pR_1Tek#`Im5lT{s-aQ$qzCvMBQB*&`jhqFYXQ#Vd^XW1q5r$yqv61d(aw!%IBh
zAi^S52xgu_^ek&Y+4cDYhAnB=MQ6WRld-wF88c8y@;@
zxoT~}@Du~KawZeXzwaI3x6Y7Dn+%(&AUWVV1t|%&rN6jYdXUZpaiJQvrf(=FnXc8FM9#2{ImnV>a3l3;;`!HT=hlg73m_Iyj
z-i0=QBv1jn-(6i^C}+pMC6i=u=V?3-FTL2lDkdGOi&dr1$Yh62#d;TO@N%A%+w7|*2+QLu`P4B?cLLE{
z_~Q-s4Xsaq+Y6s$I66KhLEvYQmp-GQH7A-iL`bHGIuvrOx>%vau9&N1lS}Z!b!dg)
z`>135?y=S>A%U{qA>Mf;S6|O4G57xZ3rrehy|^
z3PTB?*cxa^ZN7C#N#typ)xXy#RpR7VT8}3kuov7NYr4_6;lS^Bn(}ZyEZ8@=f@B1L
zl1Wge7)xu&f&5~0KfJZ+!Y!F;wPk<_^%P|lH>^&)^$AfT+HQm2lI&*~{Dqy?P5)3q
zGRBgB`Un}Maq=e^%lFlo4d%RWDw?Za%6@?9g9R%>Jv!FLEAL`l+07`}5(K{ci;e?o
ze~$pRhm*TXMnn20gtIu_hW)Gym@s2GO_#{J&>z@lk9RNjb#i3orgUZd;w;}j!K=m{
z25rV7qRk`m{nU^%VD{fjRxrv1881zpCF+KM#eQO#?N_OP1O*v3N;WeVzNMaoqD+EW
zjr7DG8!3!~9CNJkWo55VF)5?lB`2G<7m{R?!{Cx9yG6130md)`l5SKf`yF
z1~V4@4*eYwkca)tkca?YLATjDvN0O5QUO8tkp~Zk4%=KX1CErmf$_88SQ<^Yv!ud*
z@PTrXf{kSzo9YCjS}F6$B{NTX`YyP>c~hI5X5^L;=+<{Mq+C1GvXN&o2y^%F?!;Q$
zoc)9ijoIG^HL6YF`!27&|QP6-Kl`JK^!zM&b$&AYPZi%1a7)z!ohw~=^aRA0INsG+7+9aAwteW${|+Uj5;N$#U)?B}&)Zt@no73BK8`1`bf%ED@HtM2{EY
zlb7GfPgCU%R^)~mmJ{$QKn1>k2xBlQvnEtUtLu?87TP}QHt9PXMWyYssqyUb9()bF
zyb1pi-H_n9rwZqC*VL8K7HTv_>($!(R9%XQs?Ih-wa8;R?u^@04m@(t!%s=oC~G5O
z@Fg0telfW9wnD~nXbUFr>$NCnL3vE8U@QG0d94Q*aD^h-mFo8hL$=y~p&0W@xB31h
z@$$j3abDy>5sw-kod1@p%#{_G)+Ef9qQ6;4YCvnBrozLJcsxc-v!~icMkKFOHW&|M
zgAJZUilNy#HG>8Z+KTH!J8Mt2Nf_7+**xyMd;LUJo^H5u9ZrmBatTrHNKlRr;iB*7gkunJ1>3}sYJjay-SF}ke>BwDwF1aGt5JdM7WE#1;
zmZW}Z%$)lfBtm9qQ0NZm+oI*fSL{ANh>2l7rhNRAL7Z_-XP3NQ5a9hWW2{ekylK&i+`v$U1gMU;DCKBl?zQIraU#bakl@3K`H
zaVGF?XZ<=tx=OUn*a_b##F;(&TF^oynM~-sUn>6j%@tyfNX$PLajQ83qfCtw4NPj2
zwRNC*-0aLlMsZ;Vf+&&J;WUR^%+~5I`T~Ib6(wSC{DTXB1N^4spoL8M(h)Pa^nPZ)
z9d>=<8wu05FxHBQ+?NI|1tS=-svGW~{dT3_sPx&pQA?O8(XEr
z4=P*-J6S}3GDO^QUy%gUd7;*L0p4Oo*Nb8DKi|p4%zl@B&AU-eS@ePXElUKq3p#qD
zt3e8XkW4SJFC(nEK%3Xa>`E|2?IoPodgYPZOsNBaGeI8TAf`>|D!GMv6TXziT098F
z#y^{i5~5CNob;?rWg2$5juM(w(2NTA4*DGGra2#fR;N(kaLF6fRPjAq?ozTznpzR&
z(VL-D&CtH65YhxxeK@QcBndTl+q1L$P+}zrRoc>(y-=o5;rk(Z#arB3>o*|7e1c9Y
z{i{cK#pLt@)jMh_$URb;-zZnF-;;tH2ZuUPb?vK|-2kr9w??Db_4J}xNM@K*^$6)_
z<2J;9i6TS}@Z_u64%=ji#Y`K!u6Vtq1v_C+MLu^Mcm}T6pPx7uQ<-lUT-*127#aA<
zVbJVDhGoT24Kk&E@$M2!x|qrA3jeI%u47wf?uT_y*yJgq+3yP6eZ8erR)jI$GKd=i
zX8oQn3`by2ZYwgot@OiOEYxb@EO+1je|#2yC~qCBD&mYmUU
ziH+`5Ifa$yx~g1?Co6DjkFqrkY|o_QgELU>PVLf%&vWb=K--i@cm37PI&7|}ktlxb
zno7Ol>n6{j5;awN={YwZZ$7Nq=NRg858)CXu>rT^v0vS3h0Kpf4F~!DHfQ0`^RaV(
zGNKDK!NY@lm%qrOxjOw$iu^<>*b@3x8reUBp~9gj!E!F>VrWBn!5hlY<5B8&SDFuX
zZnJ^pD1CUvFF_^EQn+Z|bK%}0g{TgPFUXLXqk-R7PGsIdzu<^n&V|Ssfm2F7u~)hj
z`>gXf-ak$2#*IBSHUCJ}*WMk8c{tsFQ+^b1UqFOIy(5*7rR!dkiRmU2M&+Xr)$Awr
zWL5p#VVJ_OIR%kiCOT7or}6+i$h-rO&nnYtk;eu
zHGgzUB_)^AY5D~fJJK98N=)6zJ%Da5#4YI4*)45)ih5coZ;G^*05|NG>3f2I&8bS`
z5t(vyygkYE>D$TgC@s9OuCj??@T1aEi_QEK=rUGep6+z5{n$2^*Gl?hvQ$ozW!0NI
zk@*+5iXazm2I)d|95@!4GDHpnL2gWQL<0F4W8N&ua#1z}mtIU_;hPdxhXE`9+QyT$
z|Hq%y-dd&Khg#t?oS1{R@`M(Df&(;>6
zUI*)s$XJ?+7JDf`+O5&}_CT1I(Sx2dm{wEw0C|jp#7r|QCwt|$>IzHcm_j?DqAgTh
z+Gu2OwMjNW6?)WsD?eaAk~h4PwYzh|wM8FdF)dB%=a2?F#X$eQ9m2$a^doRK;p|M6
zwO}I;bmq^)S<~2nS?7}kmc>y+(+jli$w@jc)p8PNaG0N$%#qV)&?Myw=CMs$y1*Vh
zIX{gsw$qZSH_GolPSW3oIlgK7AUM~-70ExexOYmnRuR>!mA>&T-^jgrI$mo_$e
zm_7J8R?FQtl(c{P`G7C<;S59|+mu(@;(DWxx-c*KmXZnavlT1OuP*=F4R6Zj5i#V^
zx*wE}2ih}nX!9(8Plr#3N!faDirNPX
z@9ryKV<7jtW;M68>5oH@Y~YTHi1Qc_xDT0;20H;mN+q?HV9b4|q5Z3>Bb#Idr&?#G=Nv$MS%^!Gi^*q(FM;B=uCHSZ
z{j%!$Y3zMM7lwL&@jad#vIn;*z2B{(Gba&@qvP;JT|v+{rb1-8iY}j;p`jxSlhs47
z^b7io|LoyZRzJzxqfzIt2;64JDW#>^!l3v9i!SA(5^+<3NrsN!@x_}_E`7I|+K-(m
z5$PLjB#BPimo$!UqE@R!cphPi3@h#zO0EGed@0{EMD_W9&>z3;T9j9mHk}*r*HRy!
zTw&-jzb?1R#;YX1$*d$;4WJUfz*fiHD(Q*{;KFY2zRc0#pP53N9njCq#DEOrnWAi4
zl~IwK{g&Hi{z>mkfqE>WJglMQjoX!|jK)y)o|qhct7iwH&12tEJIQXbCu6O4dj9a!
zLcz@!jW6tfp!%~4?rE<%CA(`ha|oj1vWKw?CB4p|x7}3U#@CT1*Ts~cs=TQLe=w<$
z(@a&s1J{pj9)XdiDuKm@n4<5>!1)%}U)r2*`>H-HG;))dVNI+)-Rtd5ol>z?BSHvQ
zs~jEu)`F{7&eJgm;-lr~@Fg&P1$n(UX_2#xF^I!|##QHO?oRx@=-^&HXgAZMA>2*k
z57kzgPJ9E#DYwS`e%9E0Qk589MS9I!XtJ0b{${jeE0VdB0U3?bfZp#I&ZbJ??g+CZ
z@R{%LFklx6-QgerFSK@N&l%ESD<9#~;p@VCYZrOMZVi8Juz4|#Q!5O(CV+LyZhzK-
z8D*z`N!=PI%IdLux3bt!!8R_BhyS`U39J8`9Q|6nbG(ji6;DW3?5D4!sjS{;5-BN-
zSx69O_vO$RS8?Ania`^9NtzpHBff%&P9a)PA4K2)W;D8BweshGJn8$8tT8Zhe3F)5
zPz7!QS61)PS6~yQmpR}3p?Ja0e5084cGqM~=9%d=CQ~9Vqs+2UhK!
zX{I*HDsXIS<9+oj-3!3Otz?G$yzlP2*%mxQMn6?o!JYb@d>c~3HWBGFJq0x`)eoNO
zS=Q|Aq?Ud#?Pj|cA5wfUXmOb5x%e_pxp|rvR4iL;H}7%on^sdyANJ{&SvKi=b!4%B
zr!evGPqSF`%dyohZe*`E_}VTdZs=URF2Q|bC!vX4$4ih0;Pj{q
zHRxr$_D~wY*bi=)^6NgF!4Gt>#CRV;?9D`F(i%S6qeY)xGaVwm{ajMP5?is&FZ
zL;}whcEc@JrinXcQ1jsDAy#pUipS1>J4%_Uv#w6Y)UBX7+M^Tl7}DWf{0k}}+X48OWc
zl3m;l`n1jCX?Wq@BY#cF8M2oR+1R<9CpZa+3isIRfk51QD!r~74Z`{X6AG@Uc^`qj
zh0CMgc@x%=(kMU9kJYWrP5SMOBmXYwVKKbq7GlTg(W0MBIq=Q+7|t_p3Gy+=!9r-
zhylqqg7T7iko{swdsSJ*a$xY??0HJC$z1wntNb;sy|)9CK$@T>BS<{0y_zwAjA0H*
zsx?o=(Xbqj6CKe3bN~q1|ZT4$T;3-5Fr|CPRoFUbJ7ib9sjk6k(
zqdZo+Hn6K?4_q)8tO>g*AO$PL$v(Fo*
z_WSRZ!sOo>+`L9beJjJzJN2U;Y-lPXU2+wn%_%Ibtw|2$7np|RiBYVyKdSYs=$B2B
z_p4ZTrheNjR94b@E#V%2(+g>*_uB8yD2eb@L*pUUEV6@wvuAEL#&T{*_Ax^;Z_vx$>8
ze^+wKUf_%NsQi>er(GtYYi)?e|B4{f2-AjZDsH0`BPl;@aJ=4>;*-0EPl%7nu2BeaRvgC
zSSF%T4GEigUu)hL@}qo`Yh@G^b2;V~anK~F^8vkJRhS=oW11o5u7co>QCa8`I7fkB
zyJpp^evqaU%tn-dCY3O3N*($Mt8X_S19a52N@qmX_2#5~pXZz0^g9vLTlZ=C@tB=li6-{)U5!;ES!~*8WPy{xW0D5
zu8x>{K<X}v8TI##Q~@mZ?`5&-*ED!w1mDaNkC|-&5_*WMaoG}ClEj~xW~rrFo`xzhwPV@
zPqHiEcj!U*><5H)@@y08q~ekL7B^!N+UF?HZ~0RbZ3W@W5MnXy|XJ?gKhO@bg`DPV>_u-lcp
zdw4Ba36@3(xlcT!VB6`rh~5EPp};;$R^6b38E$7B&hIS*4c#y6bi7YA0eN_-zy;Mq
z2m7Uer+9Gew_n1icpsZ%ly(IDrY(WL&SYh4@?}O-N7X^wq1zSi*+r?LE{3)cdmEqp
zMtf}9Z%rHpN^6xdEcYjVVZEu%Aw{=k-0o_ZOsL*qU`Px7a+*tj
zRQK!|Q1zBBr-t0N<7ea^?ui-m{02dl#7eE@4)=QT3Q;9lE1IS2(dgG|?XiMSfor;w
zHPnqd%Z*$TZt2jl#vXvlqFej3nHkLRXD={h10uVbjTT=g#fmAvyLp~n&aQOm=|Ohb
zSRpoP)@}W6CnQm{GpFv-hMF;bt3|kfEcrrnpJuX&TK@F=D^E@3q%?ixcUXnIb88P2
z1C?k{@P+IoLk}Af{WP4NJF{^B_wl!`XS
zXr{@*%Ntz<#Yxpd(}8*n=7Wf`&fNx+&lvl8ht8@KI4@YOyKKdef%NF&iG&8J0zCGV
zA7q2$Jod1p$}4qGpa^;>{?vni34=%MM(wJlLJ108hU`R+id+j7$iQNHXqHrG9qcQa
zpZZ2_a%Fj0SRN|z6!}yHexe|~=pG9vpqc<3bk$B=Czu+(x1Q1V{jSP?_F{Hv(s=0}
zr@|mx`=P7t^$nFAG~D`*bbi1h8>OPl0~x_<5=w8W0sV_F5v2i^>kC7F9QNzH+U-W~
z2%uJt3{ahly=SSWIyjqmH%r9pI^l3tSS$#^M^~sjq|!>hxbB)^N8(IbR+*l;k
zV`nQqk=8GxbbIfbUS`S;kb6@`>uu~Xxdra
zcNG%%4qk-v&*OT&Z2*caZQJg%%?Vdece5S_QTdg4TjI&?GJ6Mq1?j8Xu|ztl4ELg3
zCg%rrBG<@1r8yLWPPBfoIkvKSbl5jx3S>>d2`zqrNKsISc+_x%8~-9UZSyI
zS^4$q5CRqlu=;(`zvdp)iOiZbBJd^3X3P5%bFEfriZ8Qgr#2EI&AgA>E4yFEEBaAk%x
zJnEZ4HHfyeD(r=^e_}XW|~*~RO3CJLRS~I2)R@AcKJmo7W~<2Uq{`%+6qC^T*w!bvtUAB
z5Uw;87*Z}+(A?T2&-f*2sZD4vJQ8nlU4wA-tB5l_5KPy@h*mJMht5$;G6$}~c(kn2
z>7*wyIOH8hm=bm-T(}r84j?^iYzzY(Ol(xKZ$SZn1ZK9@V~LeCZVG;i=}z54xA}fQ
z2|E7NoB;Ae?)+FYidQm6aj%!Xw3W0*OOeMvQ3ZPLE=%>zNTV$}zkB1-cGkol*>o3l
z_tfQ$DbNm1LvrI&mc~0<-|ha+sjRk|m+9^PGB`nqB=dK5J;QNwTcwfntx1fI3dQ&?
zeut=k#$S!!A=tWvI${F}o<+s&r_y15%THUka$PBj>Bq$lC*r$By(OOjuhzurXicva
zqLWm|oevC%ny+X*T4|T0&?=>S58*~6**Xi1w1f?9-)6*N4)rRhw)N`IqAT!v1`+l^
zoY-9>^*?3%zJt^^C51{ocLBmQ0avWyPSDhU8Dy?nbD4n$Fhht8EhnoL_?6j9uAYM)
z#PcuyN6j*!-4dolm6)W`E&6^e20olYiSZS+Xm^mcqP(}vH`*Z_A$}wMu?(7!RtLtZ{pnf%#mQ6~=Y@=-iuEq28IOUbT#@$mn5Q_7`qA`D
zPD_R|H6C6Wz){U(4G;9AFl&Q+L!k!Au3+t4R
zsNCZxC=|=;qE8PS3}s8fRuPY`m$wML)9kuN3m>ebSK*yC5!wSPXeZmyTm_h9$6!B$
zk;{)@VqIzrOr3rG%IS_}azP{+qD8NlC}IOe6(_qSQ)u9MNUG%YBjBdC57J?O`c0=N
z#ijj?WCLXB(jv7aihk0YVms-b+F;rU<$e^RRlKwX8bB5OLcpjP;{v^9dJEq3^aB>F
z&Qg3G)`{_>V5+&EV2CYYQ@6XR{fS(se&wcbz3mW~g?`%BsSKrCbSyWcVIJEXkFkvu
z0%pBQ@q;p0Wyu_BHM8Gv;k9;uK~wW15dBeVIN%+reo1791$CJvez|ccYnPG~nJlw9
z{oZ_i7$`sT%a!%SQ$LDw3R}E(6VuYSmf>04_A$7~M|@vbylrn{%kT9v_Mm*)Pw&8;
z4dT1ft_0#bK9)1b$Gas^8E(O42T3FXCkZ0q=z4bIn6&b7NyLWeyP)lV@p@k$oMXC^
zpmA_6G6qyj*d0U^=re0yFf-4R)s5ZZY`J^$iVdM?{X{riPvmzq7HJ5cH(OT(RjEWD
z^2D`m*{JNHuLgurn=W>0^pr`@M{DscgF@iOn4G6JaVT+Bwip?$M@6T@IGw9#*FL^E
z5!N=ld((LU()tBdd-5fJ6#a&Re9;IhlA!Yo(D6`!(ed@Cl}w9f%)D)S!4bFX^(mTQ
z#yV245m?kJ2Wpk87L5`N&WoZ2+C~H$lUzaZ;x&pv+{^W?`oiXMT&`iE7!K>v5-||n
zr!?5MDGg-2dlO3fr7c-%nR0OzX!_L=j3pxK@*s;^oB
z13tcE&5+&f@EtdQ-FO937h4!BY@6XOOr6S^9iC$_Y(Y6wSSy^tOZgVZsjt1?sXeQC
z`{zP<(X^28FI_;LjvIE`;TQ;~^AvdttJKgQ*(r(9Zy}bF?k&K7-eTu*AS(;S4|L~D
zBsYO8Cs7S5rZ3y0oKA#RG1La`1os>qYpO>_XtwIbP$)AgNrrQYSE4=W-uLX%18)f-7@LVtxKqX*;7(<;rA)&h-?N>)-X
zF?u>mPn+e+IH(;T;+pc)a(t5Km6C{Pg%Xl77ZZ7B-}no
zRs=_X$ug0O-l#e042GuLRo(wz0H7;h)I~2DZHpjo(q_V>9$lZ|t<#aGa*a{3ijMxD
z=>|cPdw|j&`~p$&gX?1+U%*(hDm4j$o0--%2r^s0bW;
zk_`*P)Q3sCH5xgoj8e-8PuxVtTv_nvydL0xIB(s6GBVlB7K5G(t247Hp
zX}?}z$y#r$Ecw=O@&B&VfVut9e8q_Nm2Rau_|wobqBQf24?9Npa<%1YkH_1uQQRzl
zzs?7Nk)8t{&MNc=_Nvm}i*#P2&jb{S07P+}hbduU&
zs{9$NSa>iM__s)XE|N6H@T4Nk9ZlqadSXLbn|bEItFmig447o;X%gJbJkV$2sh_%l
z@Zv!z;DkHio4lFrnZH%7te?IKQmQbJ8s+S2yohg8$M(&oH#OUz$)J%b~{U4}o+i4ra**bh{l;FZ%~M`Ty!))qL||gGX7IPAPF7B3CT~i{|9abKwo&
z-MeO2T&a@lhx#mOAWB9EIf+{|Iy;Y+ixfb{k+J%5cHqvECk=-!uUaD
z^x+1R0-WH5tK4_!>%078VUZHw?_Xz!AFY`rNB3;;Z-vmHK|r>PwvOh1EV3tBA)x&)
z?VXcKo0k@EESZ^gUHh=(YKd5EB&U6PbQ)Nu&b6=39@60$ty>NY{k)jF2?!PE;ab^pal~NWnfgEwkEEcBKENQtq
zW-QZpFsv282@{ERrhcG*7vntpAjIp+2{nqF^j=cPwGK_!9;tCjoI%ZgIK`k&%wOur
zZ>CxP@(o8Dd?JZh=)bJCA^7yF-GyA7tJSB|;!m5EJcD#o<0DKWa9SE$Pa6Hg1l{@z0;)
zAEmiCN$Dl-LETbqFc5)sAkmzhgP?SE`x>;M@z({cLiBJo%I!pVX5L4>4!VxfH321B+dYTJ|Zs*BOItyS$0MiCbjHG
z-Qh}RX+p6M75G!khT6eruJ1(zZ0)6@$Vu0
zOBS7}HYc-!$)P?YUmJ`lfxxvb7R!n|T^WaLg91++H_=(p8~;TS&BPL@g(ZOzcb_q}
z@mpFU`&YDoMa(;%ow%7kPjml~&*XUWUO7zBW?+yfYyu7W^YyAhfO?Y_HrotcNULpi
zsU*TX&F6eYj?;bCMz$x9r;HHNWT8~-Fg5%<@1}q@jr6tVYEF`%JnayP_Kl%4*E
zmz4Ya=T-xtp%=kHo%t!!p07lZr|Ghe13D1Ww
z$yP^|SvqoQeYb2sb`jRT$eJC0>aJ3}9qn8iU>nl@*iFMHGm_o=)Xv?RR91E8LmGFJ
zmIQ_BwR5r*Ii4uTM4|oXjhdt?3Fv)G3+^%GGr(Wj9{_Pcj=vk2PWhQ@&CnyfU=}+m
zysBdq)>J3)e}|mtmwLx(J|T!r^rJC2aAFl`;QfFh?!Os14ohQcWmDs_ggVAie{VzP!hrlN*>cHjH(VhjljK5$
zN93aukd#?MJ4}Y0e~`(Y9|W2`EwEJ_D3l^u8lk+I*_5*3;-g|u2$JBsH2|;ZhOc&ooh?x}^x&2H+!Y};
zMc3;ge=O3NLE&~1E$5=F1ucI_z26aPwn8*g<_Ks67C5k;Ot#Upv*;
z9^Anu=|kad0**$#sSDiEyBAH)_!@dljh^g;f8_>x=8t`^b4}Qas4~R9jJ=lO(fBu@g;jD;S_OivxfH9)Ye=naF;Pt&ZmQa2~5?vp|NR3
z!F9+{UZ1k1F)?i;?iis-t^|0x?Fo9;e{haG>s9#|Sl`6KI=K44Dk)qf
zeg@fiktfpH=LM4fKr!}Ri7SJn{Wk8-WKRW*0Tf4J%cXX~s<=)Fg?)ov!-I&ge>ibg
zFupT6bcJ?&Q)8%?;{DgNH{;J-uuewe2mm7}5gSeQ!p%JPJs$cECWHNSiKE~y8twXCF3~LgH)+G1Wj3&X<
zGMOsOmH7e9eWQ=6M0+u&e_gZQMVS(nwZ8lD>t^EV5MfUY0+VT3;?4ZasMZQQva-vQ
zX>d&FUtZGsgb8h371@&CNh7)5vi%~>_i(#+&Y?7wGHEKVGj5
zokjnFE5uuc^oR^Q*oY}0vz6_pQHSzPYZRVY_d*$+#x)V*NM#Uw40($Ku-@@)3rs`H
z;|LFD%N6^=I7!~ujQ&d7CI=2i_vBK(%c1<{-uN#Py7KBOxiT3j
zSIy_gn_BicAwXwJ7{rR>`Hdnweqj=LXL>dLDPPzwKmJ!efBW+$Ke5Wti8Wfg$Iy^2
zUyGua0|kR_i%a{JI_-I;$4VjnaeLjdeZyP=tR&#u?1LW_E`IX{3~rlfvO_b8%g{I
z#x)-_JMAC!f9V6Tq|@UaNhJN1k+|&dWM+Pp4D(3bo2PAOt+H58f(}4e$ARC{aPC2u
z1@|vJxeF*?!mM&$v3hZIm8PuU^7DwFs1y<&D_#6|GUEKE3z{H4Yqco5u?N%jp6?%|
zEZpqkaF)0wTBIMuyds-wk4tl&5lbvd)_cy}7*c3#f3QK~Kg}jKLi%0xo_{*^ul|#t
zR{0hysJvlve!*XETc(EVf|a6za9i;~N~nGle+Cid4i5J3*%$68>42bOSPP$_9>_ar
zV(we^?aw3^O*QF@@LGOdSV$m(I#_aND7UTrK;5D9SSieKin%CBqhX;Ekah
zkv6{jw()+A>28u00{a>8!XFAr2qE01S3M#9
z3j>y$r5n1jv|w5Q)?VwW6r~lue{}yLy+1-Ut3G}K)zu=!>D$G{%^pG4GiINY1bxjw
zMeB4cY4Vy5aj3WTkdpyAe0BPzFlONdN(Dd%J)~7nuA~>_D=SWZ6jZHo
zPjp`+Q$(92%C_ni0v`v*!e~wl%xS85&f#e}TTEu-#CVOGcl~QL9CiB?7k~@o`N!j(
z>xz=@2o?aJa_Yarf*x{$mJWSxw(g+Ho*x7XoqauX-aBIo!Rh(mo{=f&JD7_Yf2q7o
zM1ckeyQHaZOk#~i4t^UzB?zA11<)YGSjY5-b&^3xuf8sA87b@R*A~Xt?il*Ka4PHc
zgW`EtOrp@%g~!+RVo0fn?r@n0V0`@Z9vv1Pt534ttg;$<6PX>uW(v6{L+l;}0&*M<
zhvmq=SWsMiow?Hz8znLXp{)Q?e=ua+^e}7d{=nWe1lTE?e=4ea1^gb-
z3CecNa0_lmJgHm8H>fH|rlLFW9}uW7$&5iC&!xz+TxXV`$pyGCynANt12o@f#hG3Q
zv%*ew!9uBU_B+7=N(mjwRR%+VYc#@H=P|0J%I*^ZLkc1;&Aw9Gs?K
z-l=}S`~kGE<{9~)a}pF#e@OvY35=UicslVcV%;8$9|EE~_e&oOc=azm-iMd;_pwPfMhxX#tdr
z*(E|TIu9{-xIA3J*~oq$n1=za*RFSLqOO!o-rJ0@NIkmXe>OMFOfcUj+Dtz3)O|ap
zdIk@<#s|ff(3LoNjXVAJLoUD;Xak2D-r~y_Qx$fw#T~qSben;OiiSxP*Vr6ZS4xUE
zC|c=ao-Ylvdp@^ZO|y@x?yVo%pd53s$JN~VOB|=cV(4&RZMW|+4+{}`FE;-(%qoSk
z13WRI%c6{~f5qlB>&F#=kP>&HO>B1r!Q38U*8A{@Ol6v3@CC;F;h^bo17K%}d2;{H
z7_Mt7NK>e2AdmGLg6+e$k`t4>On%_NpE#fRD5`aZ`M3J|afqO>PxzTs_0Y=+Z12NY
zVCnLc2wb&J%Zrmac|z?42xdMwxBmrIY~6paf7l9?e`Rzh+jSQL|Bij%ZUdeNKSwHi
zooCqiD2y$Y;v(vAMOwxC!kKstx}lqj+UrWniY?&HWB&>4W0Z#Gu5^0Qj_K%BD6H76
zr?Uc51FE%?=1HWy?V~ETrGm*JJZ`$eL
zVbTB3e`0^Ey0lA+f!Qx(NF2`l^d?T&kPR2vVMc#x>4}mcO@)q+mMoaop514rnV_QC
zqo{Jxdi%T=sXcM?qthW!2^6(+gM9*E3s|T>yp6clRKT@uQA|&H9ZJ}%MDCQ)&fgv_
zDZ4|4O0~Cep-h;hGhXAroDt&q!|GFHUM@1fe}_xVY%NpsD>JpBg5*<4TQVvUtli+1
z)G&9@CDk|;+}vke>{#x#)vi)>!Xw{Bns;Gp+9nfTV6F}a)Q(apn3ENT{59=cgqCIH
z{J?VZ=@CqIbBx-Ghw>9y%StuJ4)?Po)7sxc)G;Ja3U!}p)(-atc(vjzdFan{PYQ=M
zf5RkoP&tk&VD%Z<9$UH%+CM*VVZl^jD|
z%_{{G;eGe3I=J(Ip`Tye-1}&8Al)nFFjHdy)y>c2Vpe|Cd-+Ra)_Rc>k{YZdf2$Z}
zlw9Iv@f2M<6ezh338L%CD5vyXDPg#=?F0&u;N$)FSemqiY7^)xdMNhNejv!L^i94>
z{uwT;F5IvD?JudJVo2TR@F@ZlUmMiz`NOSi#
zP=O5k+jhUqEPY|%9B_o=pcWkz=C+bOM=Y~s?7Swj-7QgP*QCe9PZK86B@afJj_ZOX
zSv%gG$LwPDGKe=qu7f1o?#6@51*x_5w#v%(@QfT5q3*ok$WnG<%ntzLe_zPgNZ|(@
z6v&JhW-cM~ow*D;$*mD2e0X-GfqH~$Tu?$#Wn4aVYgutnue6y?iwodo#6WgsfyK>R
zNWOst&924WWg16^JDfaiDs*^=p0h4r^M@>zg+qby(9m%1Hcr!1bpUaFy;r13*SaXV
zi}d#?y*3kmgH8ob_u}F}e~=6+$Gz{5xztxin7oZnH%+VPIRv{p#8y2%ca98}Sq&{s
zq-Q~|tuczi)$4;2ee_zoxB_j@It=(v9?vM=5*(M57Fm4@vV}m+wDXPiaO1J;*F|m5
ztLWcjzd6*ng8${$;{~7`Vgl0mlFy6%5v*C?^K2!msK;@rJ)FCHe{F@Km5912M88mI
z()PJB8rc-;PJ*#|23%q*Dwi4w8E)qX2VNnI6k8D>yX*SNns(HwP8oXgmsKwqkSJ9{gPy;O)h9(l(ZKrwA4Ib6#sB7um
z9PU5M1G%8&yS_GLe=Lq&tI=v7c*lLp5_d9-de5zN+w?+~e}ar@FbGCO*7(LY
zWrxj%PV88*d-G8RAy|xyv1nk>x~sG?Yr%|55h`|1w6|$X<+=#g$_|MP^e4m+$6n_T
zC3XHf4ev28~jl
zAR9&9M>)k5MJoEEmAcCv$PVJrkz@pMb3_HEN*el2*0DasX*E+(mdYvmA#QnqW?D6Z
zGs85DYwC-qsG`~vL;=vOsp2_Ao%E(($qS$GDYj*sK)*fNtu33*c6b{Ga^{5_8LvE+
z*tmv78ygawe-J$Bah4QR#~ctO)J}>7yggY!urHp)q<5HK{FctTRK02c^{J<^hfBs=
zA4Z2`C@aJ_!#U}CTwcywi953cAFS?ja2m}bU2AO7O$9NGB^}&}(NWiIQJ4Q9N(kDgrXmy8KlVQPJ=2AbQt>cHe>nAqUQ`T;^
zp6W_o;v+%#>5FXFeeS}M0>}fYf0YR5KDUcxc>`Ui%;jb)a&M`@uVB$*fE_8D&{*zA
z+Y^GIe~T!U`!E&g@k#n|lL8QG7h22$F4-2{s!p)_cY%y&c{BdP#X<*Zs
z==+mGuuFYl+2Kd|(^g_(fv(
zaag~;tva@riQRouCyeCx_?tV`2;*_Gr;CY*9UgC*(U3#q?{mD5Zh-=@!bcoEjf%w+
zYS&GS2u(^0R;E+*NT~oxG2|`a^M);9e}>N%u{wty_>c$`j;spmUvmqv;NLp00NPo9
z7-(AH4AeF+08L!Q|mUtFRqkvtu4!Ok|
z)?%~8nRzPKV7|JQb(<3_6nZX~BkkOQ#MCNkzzz(RtdrS&|bUT0y_o7}fd
z7M!)iJ_jB$llf+M!k6^v{|lL#o3FGPnOe}goML=wq}*SH}UY89~A;nR~L^cbPj#YQ6g)t>K67iU5PFX#wG
zBNIlM`cf626Vqwyaf|0s8tWgKvI8Il&BRpCN^;-w=d|T!&azmx?afWO8BVbMg`t)6
zTwkP$;>f4+3ELNY&!oxs|&v$$l=pJDW#6x|}jsL6L=d
za1dxgWR8cE_h;!rUmRg1RN_2O4{LmVv_wBjnp|xSI}rXMefLA8zOZBef$9
zf3EuQEpjn6f57;?XN9%^%#o?OwgMM-D*Rs&+Ttg(dZcT3;U>Yj&qyPjrEruz{R}DV
z3jhmn0LceTF&szpP5I2N(Jl~(LB#O+{|PscJ_J@4L^~8u+>_frjPIk`dNZz)7tdb1e;Mn-+*hXnAfc;_s7@`9V{vpG
z4$ehcLV*i4J9g4>%Y0V#q)HAd?IAT?k(cYDP=1re(MAUg=xzTIgf4_g8^2t|5SoT?
z{&(9a(r8bFW=1PriLzT|U6Dg^7UOQU8aV}rL!Z%xjpRNx@rdFjT}%Rr
z#9?Qkk8M{59aJ1V%*ju_=QXs0uiD1+#i8#fhm0GN!JI#&6|dHTx3!q*oL=7Q${mJe
zDYc|n4b&@e=;mBz
zchJmV0bE;)z1wgt%B(ubvhUmSi+Cx^qKPP8l^H39#{!Hgq&Xo&lv@7l?upIFrH)_V!Oj=q?zoh1KUAiSa<1sJbxdW#ztvh~;&VxB=8T+025EQZX7e
zg5mg2Xx^kW=Z{$3G#`bIDIDbx5yMOMVl)@I=LhOX79Q~^9?pcP9b^&oemKgjl&5*l
zv`3Jn*J2aY(;uXcYJ71IMbqe^W|fVXe@f{kj2;twc3@hX8R%=CBk+VYAk$qnc0^J}
zzd~M{63Rq4FYI9YeTQ5eEAPRty$ikC$5!)3+#81lF(8P8Z{?F&V50+Z{K*Z&%5ist5(HcU-n~XKEd|UjNQ;d#AkbEqbHgB4-
z-mck+)6VIkzzr$f$k~`nMo~DRe}y7d-@fxU>`@xP^pD-MA@DbYYCIMY68kO6&Q0lz
zr=!ZU8+Gk;U{zGY!Ic0x05y;htx`vwjqT2(BhkTS4yX>sb>8Sw-Z~C^3RRHBdr6I@
z&gG-26Hd0MJH&|f68M!z1p9v4r+nWcU)hTe0>S{t`ZTPV%LvjCF@f+6fA|IS>JfWz
z;F^$Nuv%b2d9l-xmPuuRY}RqB3wv
zXL$}%9Ncxq?hbQIZ)YyLhTIRt5K<*pKBlYi=Kml=XD+uu@5_}~WU&-WmRHgHlr5I|oUmx@nXvVd9p?NmY=C*r
zxg?_tbF!cOJ)R;kjzs}O;aVJ78Nm_BD
z(*sqGZ|@<|LB|&|UAw5LqesVfF`upooMbwuPYIqQ?-|aSNWKF;0Xo9zgc1bUD2IuA
zk?yYm8r8o;bOQq%q=GajIztM0JZ)0cNqK3c4F0FfGce)&b!QgQ*}uWcV0=`TbhhOf9x*7aKk)09+>lNH?GXVRi)_AeOy
zLR3ZGaPljjWt%w5ROM3Vp*=cH4~;BKzw8DuLQY20oX`*He}*K8X5Ndo#WFC48Bv@H
zWser-YBm8f|8Wf~UcaC6>#2&`{`?<|d|4?YIKgV{u-%dngb2-fAF%I)&XwmG5N=s_
zU)grD4CAtRJpBl0`^x-H>8E*VQ!sh9L%nVcB~9xN
zcMnq|(0dDof0BBS$8*m=jK}lz0DkoJN3tl9u8(t97)li6(sCg%udyvVF&F5>Kyl5f
z#KxkX*W~&f7*Wj`QI=b*TN1n!7p_|48=3BKFF?@*s*jHF!kOO!5iQ-}!IdBm0R{R$
zzz^8|>{Z#(k+)Buv5`=Wb-<5!sj?cRhAv>|R*Q*Bf3|DV!Gh8+lG`USSC>aC?+Z4Ykc_LnC-S`{z6^F+W9Nj
zWhpb8$%m$a6MN8oZ2!gt&2|fZ1GZw`b?I(Be-3-fwsPBSVZ&52U{3rI8B{8M%^G{Dl`oV;a2XkVNKZqN4JB}iCto*ov
zQ^)b|t1%HLAlx@_Re>%@ky)(DFG~C`p#(K}f0jtDi>ftgZVS)TKnR`4d~7M<66Dh8h4%*SEh_Pl~N3j#PIn&!<;_psZr{rHI}QQuYPmxhVn{_eaA|j$}Ji
zH!ksR4!tVYl3w-24vRX$3@puJ>*PB}KkOgfN-;3=3CezfKuqH@V$eHx$b`xZ&TWV9
zf7A##s~!K-)9Dv0+fWJBn7$}scahdYk+YNIZOp2X&yz9@=Cpar9z>;j2cuqF#8d*X1+9$l{WFwcXmuRwI=82T`~!6is5~
zp;{IT;*GeG^=T|yuHdHqac
zn;WXk0*e-c=N+@uu+lC1Q%@uRpWcLEa4-*=;PTD;lUROPBrmMd4k!10e=|$^K}^9b
zrJjL5k!v=W4-mEUpKCeKu@&Gp7I^}~!aj<9(`OIcY@h)#CZMFef;OEhJDm+;e@h@q
z6Kp(sz^WmFSF4px{?tI29E76(tL1?WeSudL{g#T#H+3dHPO@S4agGoXr}=c>8uoR@
z_vXGiD0Th<>
zan?Lqb69V+)KY2q&OkkkO`vWgfA&vrr2SV+oF3u{cddkyBFDxqqQeFDq?%-$g*@iG8auq?
zTqLS_fI6be{ibCSKTj}LYtCk3LWBO$?p-UglermON)yMf&6h-;8?_ile_>nAM)KI7
z%@rLV^u7GyzQWy;_H!jg7T6f#H&Lm3|B7kl32F7}aqlEcfDWNr_`$(vk4^hzY2jpZ
zZ-QGYacn-Py9HpHI-A`xI00q_~y9GCJHwkOgq5
z+(Jc#X?o-q2_}-0j*DW@f4_$3()mg;@N2von85bb{8T;&?Ls@5K$GEGbT~0f!%iei
zR$!uDU~%J9S6XQ%e~97MuNp5j;Cj%D;pG*$=EE)SLi_{jfCAGTr<{Scs)j#7*G$j{
zVd@OZy)c!*nDR#ke?qTgxMF)U#yi0^WS`#P0&J`t5EUYB$*1?z5jUM_c4sP6Xu){<
zuHA~-r7D-fwF46ZG&q+akOCC93?Kp%Z~-v4AdmvR76LUgmmrV=6cRKuF$ynCWo~D5
zXfhx%F*GqYm(kh+6a_IdI5jhqkq9S$wRd$?)bF-8-Cfd-bVxVS-Q6HCz|hRlA&p3P
zm(m>)k^&;#-5@PUBi+2Czk80Jv+n)#U9)Ds^Vz-kv-f^xXkVylvWQzi%s?^_uqz7(
zE4v^-Nm&`_YAx>yw6`(mU}w<)S-IH*odF!Iob2pqw6v1WAfT%a1S|z~1qlLwICx#H
z0h%C3SCE4l$Qb|?5kR8_$b!HiXDHtSVCDr-2Dt*Yyc|Is00!W1payYqWibP~K-nO$
zl?@og2vw1UIC?qTSXsONvcbi|^2^dMZ3$L@0?^zJ;^AUv0|0_801B+itN;~=2b5?7
zV1R%DW*}>zy(Iu*3D5%R0kk!Lr8P7GvKp$|YMP9!P{W#Tj*bxL|A|FXQ%hU+B|u7C
zMN1k0(s>Dx)z;Md^{E8{L*uu42~g33@_+e(3jR`8mevy2(pQt_VEdH=fCJzTa(1!#
z<@Uc~qk~2U_!Amb)zTT_@Y?}^!P?c;QIL(z!^4Bs%FV@<72<5g>S+IeJ3cLI8yA2F
z#Mur2eL92eLB9**2DX5f)72XEkAYvO15mOt2Z3Eczm#Mk|1ce(rGzR$>8}3{3|a`+
zUpDRk5eK+{K>y*!8tC#{S4mAx3E%*<0lR|0K(IMf&=u(F<^nMNO@qEc7WDrF2m(mD
zIXnOIq5LnF^S|BvD|88e2sASzdv9N$$A6y~5bWmi=C9iPXWHfvu#1h0tIIzYK>$k|
zd(f}wU4EUJ4fr=rSzJY4Mp{#gMG4w_U>0Qvv^ro`S5Md9(!cD9ODPEgcm=ot9MJc#
zK9vSrNJ1POpi#S^{VJc74YW?K5N9v8|D(a}zz`3x_y5JUv;kXxSpGT%3pYnLZLp1#
z8%SR2e{`TEv_CW}kSl;40CEC=Jk71ye!2eLIlsspzsS&%_B
zK=XD1x`P0&&Tb%I@4p%UiO@Ls02VgpuF$534i4IH>+)br2teQ;GBlEZvHqtW7&uv>
z^TY@pS_=r+-V0!V0kTA6Q-Qcbn}gy1&&>UoFBvy`dljGqh~a-P?0-uG9c=8q{ulZG
zlG6qKikm?N;_Lvl|1X}6i;RsY$U@D=)!h1@BL74G$CNnO${qw@;oxRv=i=r3hok*#
z+U%ii4P8w(zn0rCKJS0=pdD*&2L`#g0JzzIvp~?w{8#jU(Bt?OJAh4JNmox*g8Bbw
zrr*NSU~`Ct4cH36$-@f(Iy(cs(Ac3J!pXw}@aBLvz6HqhcMAd7Siulis0zT*&D9rR
z32{dIb-Jgk}Kw3qnJ5{Nn)H(QJ-DXIC4by@ieCpG
z>n{j@P3FyS_@6VU_UlUeyzM-~Z4w|L4;Dy_li9;$QgJ
zjRgRKJVEAY3$qaOSHZTa4I!m|()r_Mh_nKKtOpfDB6*tvze*e1|V*^dlMT6`#5hgCZ$Q+Yn1teuK+
zS0Ou3znFV6vn9TT5}-Y_Cn?bSeI&zHCi@S#C~KqGr%oX(=8sM+
z8Hzv9!TbXk1@hlmM&yOi@L^z2D5_aE+SQ`Yz8;~ZS&{_kMoOmKad@4C+p3ymJ!Fz_jGSn2U-XkS-FALhgTh
zZMi-mZ^-{;OC~
z?bk~h0Rh{?{?r+n-fJ3Ms-Wj#kD0mL%m@m*`9fPM4Ge9`rE4U$~+;xW6hBa5ljG*DXCp9lfsEGe>4&MY4$SE
zW?`lP1m4ImIMl?zm`~2_(S%uE=M6aO9jW?!i>mEeSj*FOG+d44nLoz*^fZ8U1B~Dp
zym2VPYt{f0zNuaHSeGIz*K^I=&DSB(pt$+bYsIZX#aVl_;}pvw31%{XvwxwL?dMu_
z3>-FB+a5}OgmK)+B!7s#<(Ff!SHM#prIC5`aXCFp*Jg?MC1>j~eQWT7KK0c*sz)^C4Z?s?bMeC$)7}d|QFqs4-5{3~Um{JRGB~isp2=|v;zANOS
z7H@?s2UwKB+-=xB(9;Pa$l%l8m9d01Ovri?o)d=$9GqiJRZ=pp=F8fY4*o@viR@(j^k1ftn8yMso!dyhz?2Pdx$5}eJqY2bbc3Xw$ds!3M@A-Ib
zS!pU8iRY4NX9b&oZAHly)>Pf+tnG7>Rx3nv+vhBLx%_akKb|_X+p!Y`5XZIC?Yug-
zy^N`)83I>|Xw|v$JYzT3e16t$tm1b!C?WP?DZS)Saerc$6H>6os^AcHte?#1?H(+
z%ChBb8=Qn2Ft7-?qv193uL#5t
zs|d-{hKq@R6vr~W8F#j@q#OwivV7@Dc!q~0yBiI}gXL!CWY{t}zZL0A{?0%!);hmX
z{v505Q^9H6z9Fb68DrD<5sS5GiIx;qMa-@;032aylMW$F
z@fr@WsOGW3RSW)(EAZ|()vTNG)z?Rj`9fbWQx?$Ye7rShREJJoYiaI}%xnYA0@^8B
z^}ML#gO1Pc?`eh?^LztS<`LhYlpivw6HSy1{6G(9A!anAS@toDfa%gb#vOot%B|}@
z6nj;FE~n^u>s=FSR`EitqW4zR)OcS%3$|U7>~ooM2%-n9hYxU$#3VxvYzTKYCzVN%
z@)^8>ADExZWq}Y2=%ATuw>8c1(!h$8aAcuEv;)bXo_gK@RKXTC4fBdf4ftNNNQU7m
zQ}Xc3sH*M{`dhirSd!ix;~M18ILxJCs08AFzSl1v9M0^Y##+V$);6M2nKWj(YBzV;
z(h}9*xs4|rRaYnH*-%Uz%fM_feZUDzbZ09Ff~=3T&&B3cpZVqNf8AS687Vqa(sblAS2HrrqU7rqW>))oyw1{^lV~Lo
z?SRB9?5sMkr(XO0c;GFaP9i~(VACc~$oCxHV8@iHT(>Plp18)xF3nrK+tdI|gaThW
zim*%Jd05BV?y+`Khlaj)(wqxDz+IMq3$N=sF3c=mL0^_ct`EjyV4AFY!wGsDi&#J9`zs=sVGR4T|Y9)IMscUb-DSdvEDRX
z=Z5Xvfj%`db{M3bGvlM=sra0sBM+YMjAru7ZqoGnZrPUEmoa!(4t)4*EpMv>x)%<@B-Y&tJH
z=D~>JseWs)S@)$x18#f5q`)0{Kc^QF#*i%c=dF|pPCn(WP?*noL>*n#U
zR;^vxjr6+NBIZPmW|nT7k-%x2cbh%L005HA?KGUWwdDiwtc5jy#WqpVaR`-g#af1d
z3-~YXi9Y(bGR!xszf+k=xsdCcD!Hh;{l;H$Y-$1nFaj9zEXH}_#(|Q-l<1^ovC*#dN_FA1i-V^yeoeh
z;t=zKk%1_d4TNQXS@61*?1<@<{vj0iq2t4g(O!4n+Xo`MMbD~W!R0n&oEWbqPCK@7
zq9*v=9ZIbJfUiRnRD=5`7#(TWZE@o83XV5ugxLDaJa@SCXiX98MgH!
zkfp-(!c@AR)47V0fCihVJ2
zY`1Nd&nc$vlgkEe`CtWEM{o`(r|CYkWgolN9LKHbdx(cxji^m
zow}7%Ko?02BB^1D{QTr2Yi`H3dd+7v5<^LUYQk4eM2B+Ek?Q@IaJ7>9k;nbY?I}oq
zlEM85kxu_=I9l;mj*`0J0|NU-6H+g#+$E8F3xth}-6tb)F*86_h)GdwlGgUp?seYW
z*-JL_Rh|cLTuP%?FL#M!RcgOei+1G<{*bM!+*YKV-G%Ka>w_^1K>nV5ik&3wRJiMZ
ztKs|N`%bo(uBHExWz9?Fa}2(gix)r40$)%qP412P*77<_VCop?iL@GJ~&8<4%s|cPL7K_BzPI0FQ_lYAw#X$-iyoXqESFhnUX4q0l(tA
zdOpYA$RU&(P`j>bXT7JNr*@kis={%9z=t7|BPtY}b(-MCatrxnR$O8gQJ>4GgXm(2
zw*5M`3GOoiYtOV!YHQ`rYs37;neH%5q{lkkig>Y*uC?`N{+Y_xwI2Mu@Ou!y=aiIA
z0Vo7C{gve6h~=NCgD0?01{KRxQ|*vt6GjWOJo&al8U_F$HPzlEQn?t|PvwTa%N8J}=k9dN6{b}~QOq-`X!=zZV
zHp$5e^w|?}50>h~J30X9C_!K~d-7MT~2{tglHNLVMZ7+Z;@X5zEKe
z)SJH6)Q#b&%-Ip8a;h}F9ui9-Q4Ap}tOB;!)$b%D>f`UX{YaT8W4s)HTpNnP0{LC3
z;)F+o&S$UZINWdqsXGE-uMBP9t7=9QOD>HlfomE27hLq(8B_$C@OVr-vXxG2H)y(K
zL>;$R)t0m)QE|vxEjFTZhm9>ZnUZ=MH{a{lTRUhX?@OaG^VbzeJ=?R0Vj*uNF_^nD
zp1VVrf`y=sJ)ZCBA3bn?{H!N6qm1cTQ#x-(_%fp3MByonjot2Fu54HTU{pfgv=vh6
zY$8WCcfWakph+^f63ARM%>3-;m4pLr-WeJZ%g65qbAfep*LOYzEj4IHn_#qHtqnCf
zo{;LVdp%VR88jfVcNPKo)P9ROYgXSN(!!xJ=FC?S%DvGAEX
zsRA3m&kuT$4a}c^9)p8Y;R_h}eDPKC`4JCMnUjd^Ff*>wkJ5ZGvWRuf!TRasJx`x#
zgj9Lg@4kke2AoQ=nyHU>G`?R%y{(pKYPxy_JE*9FC^k&K^lo`Z5y^<{qB|FJx2a}H
z?PVB#T1BVyz(*>g
zpqoPZV0nB8ugS}bq)Fbll1KC63io`}`NTO5X`JvzRl8{q{-O%v%6X#R``A1ETZqko
zP^aaiONteZ^4#a?F)5vw+*zq$3BD
z_N#|Z_n+%^*Z|7v4g>fJPB~Gc^xwq&5qP7XFC5k*BnGM5&}p-HODPLHQjYpO+sWLLEYx~`$)-(@RYu#+cj_W&4Od>SrcsD@E0b@%S_xplT1U`b5B=Qz9^oVRPFmKc!p2X03Ys2W#-%tN0|VxI3nG~UkY=ap_+%{LMqkg9
zUAoPx@+|~}T+Nj~%jj$K!j8ge>hwdG3JxdUW8AprEi9
zoigqL00>6>)=5{l-Kf=}<3>=_*Ko3ahVn8)UffC3-J(K$U&Tz~LC>Gt4runZb#|ll
zeAd1&ly{|*4lN&+Z~Y^v?|{e%)*l6bPl%1nsdF&I`0QP?*y*rTTd%?V_?=5`0~P~T
zc=0pmB_kuTta7BY5JUAG9)lv6GN87x8d;tC6=o_Hp*q#v&yK}w0lfkq+Zr=}O-1oN
z#fG9$G@GA^j|$&F#P=O_yMtdmfm&Gz9sen{BMi|Z-^$N4J_)U4_vfMO!d^44Yk(0wcm?%llhl-;A{);%DM
zh=pw1shMEz1l~{2x}>iV3)oI429`qHYaQAr=M@!kCXp7S%Qc^KGrg^Qxj@ulu8W}b
zby!*CUUc{G3b35!+V>$S`-Ivo1HjTGN;t7-h;?Zy0fTg}^2
z)W?N<1?0E=WGX2G)M7epw-Aqns
z#Sr5*DaVNU;+f`t>k^
zyYh>$Bmq6%_7XZI&Xy&J#Ew+@aM6mL7+fjyE*tbDBlQXyg2fWT1ArvN4q>fl?-xbM
z78@_&?{))~EpT6C9L0gSf*u}&pZUVoNAvK=6|aN1_6+$`-P8u3Mpx!kQ0qGp@+zQI
zxZW+knByj4mWRuB=rHV=-b;Bj+DT?9$uGyr+G7{X)#UnzE9Hs$%r#KrNA)3RaRcNu
z15&kdgy@g*o@
z>KENi@#h!@!Gp0f#=)&)>K$2%tz4IImR=4*DOoBb^%|)5*i|#$r~MZDc}PbPc1ATKe_2(MxJOWP{u!^)%KCr`$VvtMMQkRfnvZOGfQ(@FAazc0c=@C_SzHX^&QGgrJ#nrs8d1
zHThChRH@L)JS#IbBjDx*Xobh=8M$)EKezTbRVqK%9H52|5EFddpv521-e3(`*A>xD
z7SSi;tFz5&CyIeX`ekta;W)=qNg3*G%ZAl0FZ8O1%(rlT@*}A@FF4TCQ$5B-!E>6&
z4A`n{8^B0Aj{W{Abm6dk=(G-@G5?tsvC6R_wj0Zd!cq#ab85Y}eRti{lE!4SSEC+g
zrRh84?qUk%S?R2Un|~IrG#@0~o1CSBDXoF{f9&EA`Lx+~B6q
z4cj@YK#DM~uSMfkfo;o1s5bh3s2JHR9QT~%l#=^>fVcRernpy3pr&`}Mu5$NA$m0E
zrthT!-03zfpnWc>7W+I%#{|2Y;$0+90D8nIz*vfb)X|CxFyM}~fPCsb8Y=aYiEel-v2Q4c<^h9tnMC5`P|!dn5t13T=T0y=DtM0fC$;l#3<(ehC?kbd?(>u^X7;m_C7t9bw
z-+^E27g6A}7bpi3TN3`lo;taajojHRbt_<41SlKL)WQb%d(NrnUr@
zX9ZY$3=wBrRzB(>SrMqJlrCNpeS9b{^<=HC&p?#U&(@>1udTg!8K0@IHyr;&W6oH5
zfcHJaUWVn5_Fc&l4_{GaNPmb@5#VgT9-O~Q#(nQDZb@9P5)jl71#+Ibf&6yin^wi2
z9uDmb;rI=mpn)dsx`*aK*wY%deUj;%7m+zf;Rl-=C>V27h%`|byhE4aW)r1wwK0vC
zM!iS)o4;I+d%JI3kXEzPG!Ag`!Zsg0`}TJ3zijQZw8GtpJoGN0`
zUn#yDzYbr-MmsS+vN>}t496GM$zEsOZ!V*~3qsDhR?j-nZ>c+TZ3Hn(+~2iKzb7i9
zM{XwiHhF?HUIha#iJ6|axJ*{-T|eTe7)&7WvIUn-0a7v1+aO8Kl_%H|lkd>tId49C
z5Lc~2)w<5&ABkMS6Uz}<0{9?F$EiB&7Gm#CSRQY`c&3De@+51hxm!Nvz|2OLgruAv;v;cc)}fj>Ib^F681J@VL0d
z5#znAD4o<7MT}{w6q6ug(qNS%qC%O^GM3}CK^P;u_9@A(Xu(pJ75
zS19ND%lI*kJuQnq_myGJv@aV_aCEp`BzzW`t2PQAu-mb-(RDkg9+1y@B9ztBI`B=>
zNP>5_4XCGB;I*1}fjSgbzkU<>By{CB`WD6wAzRhzs`9<2)muvNRB5qct+dm+PS&0V
zS(x;~vyWew0yOl^`kaUs!J|{YUf*)L^?+I3poNUN8PU$;jP4;_kg4h@LMM2!;^1sS
z)qsNb1Zb3czJsgW@FD+G(=_OpQ2rddarih@YuurFCL0x-E1_|q2l-eRGk0z7cJ`8W
z;=9yn>CD@nNhKo6H~CDS7U*Fk!$Bbi@~Yd7ZBt~K2DH^^)2CL{x$^ESSiWQh+%*-E
z7EEEUnhc|qk~E~ya-a^5w(A?D{lOu*9uOSPM;epA;ZrRCw_$xC4f-164+y_Q)6nsy
zX_Iy^h(i+}GN!~q*e65QJX>50t-MRfv6dX*VCOkkLx-!=esPZ_gaf%5rF*bR@s%1T
z3x!0l1GaD0Q`!W5kWB_yP`Bpp`qCnz2I7;i1#>EbtUH0Brl9=&c+hw5*V-01>5}E(
ziZpN84G3pVdA*ZPNpI|Ed$P+wJ8@X4TP{7IjR;KNG8SCL4&du^k_f9YKSxlzR?lsJ
zJ;U?mq;+Hs%7b%wyX|lc@!>MgWBVKiqr9B!1G0A#;@rlft0|cL(tD7xny5PNokSX1
zy65-;=%*#36j=Ff7uX&uBg>!r(jPKwxq&|dwnOepOl9$7mE7mvS9(M?p6t&Xz0|{&
zkO8aQgN>${sbpQm)MC}l?BW&%tSY+=Y5!`c&2pHsaoOUy9^^hE
zfDL6)Y~%TZgyNxc4+X4K4jm7x?W8bbObhMx+ud@}bB~8M_FN9;EVP11vy(GE?$eFk
ztM)W{Au3U6s@fz49`Rl?`Zh8mz!!^#Yi02JAinRMIsQzlmn4
z_(Ek^?~1HM5Ph%m(KpA(d9&siFjgaMGg-Pz^aVPE?IP&3L5j56iO8B)J<9%&+Hy%u
zEP?ZrjL8NwC|sa;Tvtj$BqH=aWotxM%?*50oGg<7XsN%Kz6H|
ziKE58S&{3X
zav{+6L%L8cJEF;8=EqP6it6EQsj&Bi4V6i*X0)$&4ef&FuZo-W1NBTw>D&5BQ$
z?`IfwI69-4$lra)5akQA&iWzRi&$=iT>9BDc>*zMpjwdKPP%yY45wUg8l0MyGhpb9O`dem^MqT6Sn*0Oe2Bh;+L@w`m}v^2|h|yKr$?+u0_6avh(@#
zZhoyGL-jCJr|U!NRzrb{?MNRWe~(#_`Di|89gY*lcr?v`+t7
z4DC+;W-#>_^0p<%xKqE0>*hpW5Gq2y_4peTNdbk=U-FU)@F?e!%?hqUejS{WzE*C)
zG8iqVU5cf$`5@Mf>!HrAUEK~YX|!=URNC6=L_zDC4)JQKWy?i_O7|9hgntf{KH4z;
zc3f2_gXcMGU}%_>qhE4-wjt9Em3iuG7=wwd8#r{_$gaNMFBfv
z5I@;gyZpcm*sYxCnK%X?L(LX%f0_(|lgUBGMj^WdBEZxY*8qQfFrEk6gy9B{>Q0;K
z=vjmk0aWN{EvXR?d*zz4)o10?ezgDvv4+!ik0>AH6%_l1nxT7~(yyj|o7!+FXiwHt
zcG$Jse?{|0VNttgq~YBr6R@#gGu5ySOq;dNF)~C0wveFf{W1Gnay`IlBy5qTR}!Am
zdDrN$!`U`FIG0Lm$?ttg#q15$fry$N;hb*sKxs2}a6oCH2bterZ14z}7K<7SfaqMV
z@O_a*hPrnIi`e{56c!AlluUv2{03t#;vWU(llQ`(IxJaz6&pZa?cXc4
    VuQ>{q`nGX*CzgH6zz{hf|L}IP^2I{}jMim%`O%K*?v-1g9c2ED^-}x&tppf23G0
zgZ!f8bD4KzVCy&#fh4}aLw4oGwNRh7w&|?~_yvZte%Jh_G&ysxc$LrS(q1TU%i}!3
z2D{``j2R*C(RY%Y29|==)jpIu
zStsj8(sU_lxM(2N^3mRzQKSMFAKq5>_L3BWS!DCrVf?2Vfe{#4wK&zuhIJNHZ1yc0
zfX<}~Ohr-LOp;E^(CRaEP-cOLzDr5W;*R9yiB}BDiY$$bDDZ=t;!(i3ESgJ1UJOkT__X7&C8z=t@XiicN1U@Snv6{0BweCGB;_c9By1VQH8Jqs71D}tLffI
zc+^a|1n^%#s<{ZNv4HXN;sxAKhwO&r;#L!8^sY{$NGkB~%WsUG_TVPd#Lc
zb)MeCN7M&K_p;i*xAmREPr4uh6!f1iZHSgT-YAJ0dY@?*X1@t&DVLvTS1wGx@NFI8
zK;J|-qXy+o-+=JcGSK6jPTYU?k!Sl)v>=f*?nI4kVX7u!Pu%22rEBHba@}IQrOxjC
zqBB=?KY5Vm=_iBIhN!0E;{wV+J0MT#uhl+ieCTfhmb(CsH;}b9i}x@FP&6j{R7%OD
zm$52b?5DN^gW?ikXoX*AENR*v$vl|_v{fkGZ5OHZYGW3Vmd>!59&lZAL(4bMP%qSQ
z1cz6JP6dG7t2oNph}y!};>Hvk6~G2|GVn=3o*QnaP?4pET}ZfFynM0rH04PiWJpC7NWS*;bT<=c)AGS-&pvVf(x`{4|AQc)3MNU0t-&`^@wiNJ%vn|3tQw-k$OHsXH{CY<7UVON{##pe@%<`!*moq#{lp
z1QtF^R)q9n;)%sU^iWtNu7nSfrG@!s4}`UTK5`I!ig=dHKs2`kUh7x`)-T*?FX*(@
zDC&11Xd&nUTmbI0uNf5A8~VIsuARMa&W+3ret!KAKVu4Ll6FZk_sd|OW)Xmw4}OuC
z&okeCUVO#XGVNApAr)BZr|xyf3XL;*PW9xIwf{qj1L#f+4Esljru9Ck78-~{zIS5T
z4C5DA^4;G!{1z4fmyKJH{KE|RWPf7hdG5KA6&Bq!YO<<0S4&%)(3qLx*H-caBd@g4F=5gaP}5!1&JOP%XFcVg&#IR2;)JrM-Fo-uL1+|kR?|#
z5M8OaR?{uZozeHPa&wcbbgL__X>|9{QT#YtWt0ea1Bu4DC}`t|^OF`I%|4(y`KmG53Q
zDYC@{{eK#}S-Fr5Sl+-{=DfxO9(k5jO!1JFk-nhL4wRyinTtR-^1qh#`8mTs)GzY7L^)+Pmw$+%oQN3mk}W!A%~aEXJ3$q
zVO}Vi>G=c3RhUIDi29QV`XaEZ&LSS^;fv9BL15HDw=ngGn-Ma+2pg9&@D3P_cecb_
zc7qkNbf&iD@mESqzQQT?$;7^;;g3-L^1p9q;EzdcuVg>4ZaY{8=RTU_TQ#q!
zv9vrfW365Js8!cPcnv$3;x{7SQl5W+1ZE@!|A*be%96T=1A^WZ2xsdYrSj+$>RWK`>!inYrzeO
z0v9kF$NvO^q6Ga``Ni-*3>q8&sH{0K`~NU!S^=k^BBsFXEdLopYu%6m!;uGOXZxR-
zFgak$D8TIh4T9GCS`Wqs^DoGMmnPH=W&i`s!So-5*4i-!W(oz&@!ue*t!xWm8K%G-
z{~P3l1ANvInB#vxVzLr^Rrdb|Y5f2{h5W^mnhOHK0?fjg+7JATGW9PA1U!JVD-mKC
zxE_VvBAaEpSOOaqW*whvPV$ygzv{d}HX*f!R27a8
zDQPvLHmS-`0+JS{uoEwpxhg>l_=Q
z`Ik5AN`hD+Fc7J#3eqQK^&GIga}lFCYy1e1dnGRwrcvO=BCEb*pP-b4ySOrozmj(tl{K6cSx^{=85sq?Qgo3Y}6c1)}KJ+25V5PB1(J(+rQf0)@sT!7w
zMR%h(aD1K1fSh;cJpa%2$A%i|DkwGJA|
zH7rjcySjKNm8FDEeZ-D|pi5
zAcAt#vZ8yEWC&Yq&;$S>%!%d{P?x}tVFm|e4iKyq=jW9Rvd-}Zdom!6e!|?
zeqfWdby#VWv_@E2ldJ-c=^EsSg+5{SZ&`tnqDI*piv-?NVg-C=Z@AwVaU?KaaZ4UF
zf!L-&*t}eZVhk+84QUE8Mz+Q+KpMMyDhqp%b$!3V1_Ed?X(T{?OV!_d?fSHm?|64xQr$b+&tHWPj=S$f
zoiam~29)${pwwopA9;FiOsVU&pL;m5;o?g7C=-|+4xM6SL5#MEBYd2x&O$tvNJ`y=6Z>iVyy}p~_z@e79Nqy@E4U$~#LXfMq1G6w-8*t|`fN{C#;#q=
zE0FH6=$)v*%#m%pscfv-l|*-<*{nn@^Qda{-sY?S@CgCXn*vSZU&TlB)Xp%EWSvnF
zMh~G3m_{=W+&TWL@O<5H?+EhnWFwFp-D`IFJV*WgdGO%t&G?Q*oOb_--LD;_
z7BU7fbt3|vNbq%aqC-kSG9F_q&yNhb+PZLK=>IZ;%9Nbyuj>8s?Z3~nrN{fO*WkLN
zG(i4~dVY4k<*9%qZd|>+Jw#R$5#5`;Cer?!?MBv+*k~=aZ_%GUBM|cra-*pak!7NT*3iw=mSB)CkHX{Mm%aPAZ*+M&-i4=p{K1FA
z%CC4eAS~VyKT#X?F^*Qf?VR9x<0Xj!H_Cc6XKe2Ob7;|3N%T+B#5}61KCYZ+Dg_RF
zx~9g$pwRpmnz_V~Zib?25Ra6Da`
z!dXG{!=-;q2oU)Xg~Vn1NVbtYQ^mp5E%FTXE8cwX{_tP^^rHUIoBZS|rB~$?$i-#>
zl49`2OdTwcNNHDXl?EVJcuZO?TGf*xpR2EZBU23T4%Q`th}yNeli)QL{=6WR!$OW4
z(T#g)Nx5<)TYwfyCRtAck2PhABa7#ZAn_vYsy|6KoJWXDz9__N>o6RF(4^8;AMRNL
zbBdL()q}lJ&>-Rwp+H3uV#yy`Aa_CmG=cUgCkZK%7>laM8adxo;q+xfk472@(;2Ud
zR3}nCG9;2Dm&0SOA?2Ev@L2@+ZL%#9`}$&JFaP+mb9U>muGlFpXc)8Gtd6&8asjMw
zC|_@0pGsafy{EnR9VbcPdhI$m>24t8=YlghWJz<3@xq2i9u&{}twsZSKk5bn5Tf`@
zvwv4uZM$VY>x*k4Bmk-a$67{#`E$S~tPx54GGcW%*>@mcK%aX>_>!Q4yd^6{C8Fn9
zcd`AFADA=1kB1&KLd8esBQ>>dh
z1@*mb!k|F511djT3S2BeP$hYFyk!
zN{m7A&?-$F)4a$5w=7WuU5+phd!(8tR+9Sx;j5`&u@T(y87i)s9
z(u`k3!2W1R*q2K31?6bFRQig9iC-#H_rgj-P{Jj?k$_}nT6@k2iI@}maX|@M9u`~d
zl3=}(p<^GV|C6(=im2QRP4qbRBXm1FSYOqjk0jUePe;m23NN^#iVpE8ZMm-_Mqb7H
zff45A$PUw9^_cM?+3OfU6FmwgB9(EX~L@>EiF!w
za>{kaw>kWugq~t1&z>9sAA#06V`Plboj(H<>ryy~{H
zK_KrF7*>J1^Twp
zYTlfRVeI@SsO>Wrp~fA{l&Yz?0(Pkc$%Zoy-TIi%{eEwp$5mtT=o0Ngkx~2zArhN#
z>y=JfY#o`Btevkvh1KNz;jEm``D|P#hF(}UzuCn9EM1}k2%-f)%GDZB4ju(OT?UXg
zaXWj8Sz;p5B76(=tEGafYu4|Vaj*KjC6Mu+KyO(=eNO@HLl092qXL|IB&=`Q!7#?z*G`K^AF`mAwuO*
zrl}p-EYfBHWQav0pHt%5H0enmxZRVeZ(+=LMAR$#$Mwx8Uvu|FB}J~{oVY#Ok`T-6&c3tf)llfO7lj+YlX=#ZT4{nS`Eb-PG}YsgI>((vQ!>LkpR
zw223$Z$aohnLWy<=kUZm;qlZX>{S3d8rgC
z6TNT*tP0x=p}kV+t27e=w>920PQhdE@>JEO7_|G%z(}t1DVFEbB2p7C(L@e1q|p?g
zES+J>r9GQ}PuI0eo+lO>LT72CbmIJjV`GDX7)2q+_mDAuGt(y^$d05{YgKX17T
z%4zt_|LBt`z7aa{-q9Ll${K)r1vb?uQI1N;+o#KBqH~;TI>v=#v=?UoetZAJo}^QSp%7-cVs^iso@`
zBo;K%XQ3tdDJ`9Nq98zffJ=oLGB5qnGQv(kb~(5Q56z7zvnQ67JNCHCt^=?o@TIk%
zu-J^Lj{UyUWk;hIFpU6lV+D>PQowhcUX)R*tB6_v6K(ug_e706edHq`I~(C9Vh
zI-z4T*3X^YS#zZqJ7MU-Q_nnJZ`KiFV}4xUzTO)_Gz!gF+R{|W;5heEY4QJP@yerK
zQj9B1mGa)(yGqt-6L#;Qga%fYPTw6c;|Fe{G$B0m#x1>D^I^it5ya4o!D-IjdG2Wc
zl~(nZ>zZ9e^wm$E4W#TwzQ_fq`V36ph*TST5YK@uH~e~_&VxZeMEx({
zM6&b04d*~BO`JE5dU$XooDm_;pf4@#FXq5FJr(xvg}K-8{_FYNjun{?cq@HVi5e=#
zj0|r>R-1O_;6&*SFHZ|^Z$s~mF
zp$Z{=XJ{=n#A}tm7r+fjh!vbPI|!AYBp9BI$>EQ5!!h@i4dNlNDJJcfVzHb#EoUP|
z6jJTx!ujPve4AP14dA`2f-)kYhzw7^fzfbjG&M|-qyQ(}n@MKA=DQ#sU
zX^QD#G@>+Dg>lU6D00=A!}d4<*X&GK(o8AhONW=mDH!g*J21R=Uf}!Qn0oinc%=49S
zzn7Bpd;epoSRg3%W=GPZkRJcM22`kNp@lAyl~3VNOjyhU;Nn3vUwEp`G+eh_$#WgQ
zUgkSP#d!yUwBArkqN;wrlQ_|Tt_)6H8`2=;c5v~1^+4UU7@0GUGJCZoYtJN?Bc(s?
zb=q{iTZgq;{Ktve{EXRhpI}cN!-U9C!Jm-H?AS1K`@_(E8&5I*F)~Z~26L3g~ek2P`FYlJlywrsSyEWvG
zCs)QxDvnzsa^;f$S_9}pp*}PoxdqbTdQs7RH&_h+juu5#5A|l87tzykV||j3())Vf
z=cA=wm)!Zggc+TG4j?n4Nkn!wHW7ksQ_zWAsXVy?m|h8VzrJq1Khb&JKs>XC=haH^
z05+*R+Fit+f_B63Q&+^TiNX9Icl*PagXrxzQJqj3aY|n%`a&~XSe(5Yi-|R)!@sK~
zFN7zAEd>?3<4#o2Z;fYmz2A;Y1L+Vt`x9_3nb4%W$kDx8hj;KCnKBjEeG%6*Ny+49
zkE*Z%S;h(puTmeqLd?Bd>ON0T$CGNb(+6-Y%+*KoY+xC|+>10TS{5&5cM!FaPXJCX
z^=MdMy$IMZ*^#^R#KJ){dg$GBzg-eZqE1;E{snso1*$^bR#tt_Bs|khv!_|{j<4?1
z!4=%Vq=cp$!l7GZBZQnDpU*pwm{MIpH%~es%9P#otlqN^s-x(UNiPn5f4Nmu*|brb
zG#p#HG?9PJ%88b9JF(w6zNJ+2rVT$D8x*eaB%#Z5W+6eR6F0_Sn66g*aew)&5)(mI
z41}@Wa>PUNH?w~fmYL7Ojq$$T*W2ObV5-l;PFjtK(OGb=fSuenh)!#malzoSv->*W
zDSm6*@8{{_ZTL~#hs16|d9RH4Ckk9qi(ttoY!=-t#Y7*a!+?DA^KLM+v={evzPsN_
zArAi>L&!L9ve%v9%=Ux|Y@zVP+40$BgTDic8*6cNbasc}TiRWAB=PY5#mJN+1>%XA
zS$7e~ucI$6_81!W@?lC=qStFLxj_bSg#3p<(b{L@kxB}@5%!{S|7h-5Q#*DxE-g#8
z1kU>GAzf;1Jfz3{?{_8vpRwWgQ7q9L!Wp?5a(z)JZGF-P>OdNf5}V|5N@CM*(M$zo
z=FevF=tJf@BG-QbVhSIw9Z|V`BpVdFIcf
zoE?DAJyCSO1e2tJVh%@QvOxc{mAS=@yYV;lvB7D}*EfStyT-puM`oX}S`)P0
zt~z>nexZe8{F5L{xv}TsV|1E;^}r~(Iw{2O_HcVUvYXW-;MsKsgZ%vJE9+2v;+9uH
z3l$kjkDlKv;SSYhIdmQmR^#OX9AKY5x_fhR;>6I`P2YXtGLEJ7ko;Bn_`5G$%##~%
z7d-p^UboP}i+}z{n}OWQcm?lD`(j)(#>LbhjYdMu6yOyxoke`o^lraGkB*`0YRX`x
zk4&+0^Abgpo1!dhDrf~K>~*8ax!H*zb^rK!9Iu47^iZyZX0ogJ9HnUmXsj@YF{muK
z9p1l$X8>q*<7%ZzM3_P=y`QV{RIh4(G`+GXJ6qXCZcOG#w*L)QDPHdQEV^-oWLodN
zt&m=`Z)|E?)N~qLzSZuy-EdY)p&z6&Q>AZt7@}JG_9~?cTFxy0JH31zMq|=3EU2DM
zXGc|rtkZf|ELCjW(?k>uIM7O;2vHqKTza&TO~|9k$NDF{Sn3#4gu0SC)btQ>gu1Uo
zT^58(vK*JB+TZ#r)r!c~VgK)3cwU3vu{ov|@9s`2ne>VKdN~+of{`Q;$!7BSO{tX{
zUeBOluCv||-9~*X;Xr*m#JEjc4M%wf8Na|WlBiU##l(syOssVd81TO-N*D3T77Y|O
z))MQxv2Nlb9{$o|VpLg@x|es$n&u(9rNi!6Ld`w*5n=@HX~taa2YFnBVj^sQd$3q8
zb>ZJub|Cit=+|Glt97dC*eP{jGW2Dxm
z=DB?HW%r3`f(g?XfCY=@sXqpWN%q4hIqCk_XCV96=9BLzlm0z;pYpG?9)!M@tvj0j
zlP4zY1-k&5&rd&_M}sS5b6vVXvD-Qt0b&8}xtrkJ%vqjP!NpBMYEhZcPavDe_CqG~
zmlw^v%YVQKy%$tN-Cx=cmDPnl&PF;tl}M`|w_{dT3$qTt4Xfb?tSpVr7TUzt*vZ+^
z#K7i1mYty`Gz$}ZswFl=C1*!;{7+6>n_XQ~c)#(~?I1-4XyWf-_4&4d|d=>PR5uo6X>BP9dT$>T#
z)rXMv(QzyuuUjokMsEh0L60y-l$5r=F99lQatJxhGew3%Ac+B#25E6A{>@YNrb%YN
zY0vKQF&@|Hq7Xzp%mLCz0r(TIMI(G&Mr-ow*J(-891v4~$T(dJ>jANV8R$@u$sz+q
zr~WyTW{4)QCJ)PCxvErN%J`{#QOhRxkNx<#LxkY!q`n=r7UhNg>{V&XJOXgQ9D!Z>
zLdGG&Xi`2u2i&p1i=Pv?_C5;|5Xp3^6!;QW29`M?%#J+uUWWhT*%;hn7H+$Vra6n#
zXj7Fp&!T>vS${g4-;5A|;aXlD5jj~zye}PR-@#aUi6Zhlk4(%87e<90{S0a1MnMv>
ztf)*g|7f8SKMxzdkGNJ}^-vlKw=f?OCA1#>5Cqxi*qm!E_<0G}M2yzw_WN^Z?`d@~
z2=2x7HF9U~=zO%GgS7Aru4VKQiq|0Uc6X-7F1H8k`gZolET9+g-G#X;TkEOPnLYk>
z;qXzH##<1F&wPq`y*T=A_~sNvb_OSQXir>p1vgW=8ky@96<6ltc3W$@y7-9$t!wgr
zw-09Qb!ygG$ow0J-uLEjYueUqX``12PtF$4b_{`(7_FliJ`q>%ZxJkJp<-QGcTTQe
zjxG40X8UY+s!;8w2-kh03^r
zY~as)rpT?6f_zcD33`3kJKvJ-*5P{$_$j+#W7(Mz36Vn#HQFH3x(QQ1=YsCe^9{`+;eyQNT!^7^sDRIO71Tt-!z0k)
zG{pq559bH~Bzh*^%$1T|QH)a|c4z77u$dMGRg<9aXvb^5$i`mi=n-T?;50_5SVZIf
zIHqN~V8V8ZxUqMoPrCcT(Pu%IMeTLD3mL4MC
zn6*9VvBGU1YOzF@?8H3C!9B)jTESjuH-f#susT-1UCt(hdM+UknFhP3awAd=kKGL?CLWjpIWvOYjaoH3|Ntw_&yf%4<^^G8AyIH*qJT0x~m!;iS3
zcz=xoX09%-#|^*Q&Kl$H0p7aOH#^~cmFyKKIfa3Y^`ddZBj^(77+RWHiW-JF{uHYih*Idxz{U$5hgGlD&j+I&8grey@IT%bj1p!h
z;1SzMO2W3ndGhUbkY;;(OGQ?Xe{2XvR{vd@1CAV|*MXV(_(dz14yy&mh7~FVP1Pa~
zkau2)c-SPVrs$_Va-LL3)uq*M!p)$mrftI4=Aod~$KOp9;j51mF~C4Tm#`A58g=|m
zUiG9YFKZ9yuBwO+qg7#zFd{F<9xp;uAx>^`H2!z{h51sk8VqIyG0_d;ra)Ej!I6Ox
z%rXQXgiJga*?BOTwPu|+;J6uIDxQlCFoP$0LE^7cB#W&6A4B<{@#mQG!;uy$7^w8d&?Rr<#?YB$wI0HeTuIonRkxf}POIba$$N|SmELpYqnIH^AW;aWQ{gL`^I!XSp3|xgnTb@&sCQVfsYO7~m*sBgKJy
z3U9o$a7?KD0*#Z5P~%8lKI6ngMJ$Bcqs$G~r1a@~u`MfOBAOJY?4C*MV^VxT{Rku<
zjwCa$p;2B_0i5+9bFOM%Aknqie@;s!p1(k>oYSG~3o2ksYLb8EpA_uPRI!n6P3@XM
z0^Lz2p-urQp*|KUuq<&F0Ei1(_Q)x_Af-#qjEFTaUHKkr&CZBOsq^5H-$x!^09jZ@
z>nW<7JOwVQoL!}i6WeKWM3Fhz{v|u`p-q$5pr_26x1dKOkoPTnG#(ff*Sgwt8~(CI
zmNL|dr?nVD>4YkiSwa<39Ib@b(?pOG)8l9sFGza{5
zEjMMAhJF{xq`B}hny7--x=DD}+5iX*gw(5j1%KO^FEZSBo3;=HvwbDRpjkZm9_&$w
z#{&k1S9pVYfY2H*5v~tWmOD`wBS`CP7KHV>%UctMbKspQ@BlS?5HS+u^H(JtYIi6!
z@Nd5mdN0HtOF{Qt;C(>j721UGT8$vMK9*ncg`%fUD0d>~hv1u<{6n!Bhi;Ue8GH9f
z;_Stxe0KWW*H2*1&1?;MVB`DIorME;cj>KgC_>^VVYMYk%hd;VkBfHWc>F~(a!E)X2HZp$UvwFO)p_Y-xvAZV3G2o|9-k~BY{TZpos`l(j|4EfSwNop8&SLICgn&*|%2IDu*BGN6fYP6j(La3H`iRMzS2n@#nDrG-k>mEAd4x~1;@rWrI*S7__L*j7;wbotlNKFkBpOOFFh
zKmFc8T?eIA2PfCBLg*PUp_MVH6RZY!T&w{EnjL~~8AHJl52qxQ=)Delk5O!W&LM=C
zAU`sRw3Rp)2;nZ$_m5=4B2p12K&wjKAEr+EqQ+WGM
z@)HM%K}I8DYczTW)5CO2Sy#u`UaS%utW@2}q`Jid%`+og*+BD~b_@QE`oel|M5zPh
z_Ue6hN&N>xHQEOAt9A)KMV+l9K2d&sm#{5nW7Q*|T~(*G~Fl
zPpfu@Qsps1?ND+ULRW_3d4^e6hU(cTSxrguJK@Drvz-6dowd;Uy-Tf8lD+3_*`t^r
z`_)qRd^gwKMKHZYr}hMp?D3bg7Cp)1DNrqPV&=)cx#()sOIJtvrN2OJa>Cbhwo+=s
zSK(keq5*c2dQ=?=~cH;?Q-I?@?YQBM6#Q_=m;j3k#SEwdYZm*Ks=P#369UM
zzI(jC*w5p^EpH1jNc0pF)8rT?u~AeMBh$zjXyK1IW7pW-@4?7Fu=Gs-v792OXqj5a
xwhrPPO=1HbUF?E?fT%@}mj2&|nUk}DqqDoCi77M-8!HDR6ErEQsGJz|{{h@*nVSFr

diff --git a/doc/src/NeuralNet/NeuralNet.do.txt b/doc/src/NeuralNet/NeuralNet.do.txt
index e4f4f12f7..36df52b55 100644
--- a/doc/src/NeuralNet/NeuralNet.do.txt
+++ b/doc/src/NeuralNet/NeuralNet.do.txt
@@ -795,7 +795,7 @@ and recalling that
 z_j^{l+1} = \sum_{i=1}^{M_{l}}w_{ij}^{l+1}a_j^{l}+b_j^{l+1},
 \]
 !et
-we obtain
+with $M_l$ being the number of nodes in layer $l$, we obtain
 !bt
 \[
 \delta_j^l =\sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l),
@@ -838,13 +838,13 @@ Thereafter we compute the ouput error $\hat{\delta}^L$ by computing all
 Then we compute the back propagate error for each $l=L-1,L-2,\dots,2$ as
 !bt
 \[
-\delta_j^l =\sum_k \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
+\delta_j^l = \sum_k \delta_k^{l+1}w_{kj}^{l+1}f'(z_j^l).
 \]
 !et
 !eblock
 
 !bblock
-Finally, we update the weights and the biases using gradient descent for each $l=L-1,L-2,dots,2$ and update the weights and biases according to the rules
+Finally, we update the weights and the biases using gradient descent for each $l=L-1,L-2,\dots,2$ and update the weights and biases according to the rules
 !bt
 \[
 w_{jk}^l\leftarrow  = w_{jk}^l- \eta \delta_j^la_k^{l-1},
@@ -853,7 +853,7 @@ w_{jk}^l\leftarrow  = w_{jk}^l- \eta \delta_j^la_k^{l-1},
 
 !bt
 \[
-b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^L},
+b_j^l \leftarrow b_j^l-\eta \frac{\partial {\cal C}}{\partial b_j^l}=b_j^l-\eta \delta_j^l,
 \]
 !et
 !eblock
@@ -948,13 +948,63 @@ Again we take the negative log-likelihood to define our cost function:
   
 !bt
 \[
-\mathcal{C}(\hat{\theta}) = - \ln P(\mathcal{D} \mid \hat{\theta}).
+\mathcal{C}(\hat{\theta}) = - \log{P(\mathcal{D} \mid \hat{\theta})}.
 \]
 !et
 See the logistic regression lectures for a full definition of the cost function.
 
 The back propagation equations need now only a small change, namely the definition of a new cost function. We are thus ready to use the same equations as before!
-We leave it as an exercise in project 2 to derive these equations.
+
+!split
+===== Example: binary classification problem =====
+
+As an example of the above, relevant for project 2 as well, let us consider a binary class. As discussed in our logistic regression lectures, we defined a cost function in terms of the parameters $\beta$ as
+!bt
+\[
+\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right),
+\]
+!et
+where we had defined the logistic (sigmoid) function
+!bt
+\[
+p(y_i =1\vert x_i,\hat{\beta})=\frac{\exp{(\beta_0+\beta_1 x_i)}}{1+\exp{(\beta_0+\beta_1 x_i)}},
+\]
+!et
+and
+!bt
+\[
+p(y_i =0\vert x_i,\hat{\beta})=1-p(y_i =1\vert x_i,\hat{\beta}).
+\]
+!et
+The parameters $\hat{\beta}$ were defined using a minimization method like gradient descent or Newton-Raphson's method. 
+
+Now we replace $x_i$ with the activation $z_i^l$ for a given layer $l$ and the outputs as $y_i=a_i^l=f(z_i^l)$, with $z_i^l$ now being a function of the weights $w_{ij}^l$ and biases $b_i^l$. 
+We have then
+!bt
+\[
+a_i^l = y_i = \frac{\exp{(z_i^l)}}{1+\exp{(z_i^l)}},
+\]
+!et
+with 
+!bt
+\[
+z_i^l = \sum_{j}w_{ij}^l a_j^{l-1}+b_i^l,
+\]
+!et
+where the superscript $l-1$ indicates that these are the outputs from layer $l-1$.
+Our cost function at the final layer $l=L$ is now
+!bt
+\[
+\mathcal{C}(\hat{W}) = - \sum_{i=1}^n \left(t_i\log{a_i^L}+(i-t_i)\log{(1-a_i^L)}\right),
+\]
+!et
+where we have defined the targets $t_i$. The derivatives of the cost function with respect to the output $a_i^L$ are then easily calculated and we get
+!bt
+\[
+\frac{\partial \mathcal{C}(\hat{W})}{\partial a_i^L} = \frac{a_i^L-t_i}{a_i^L(1-a_i^L)}. 
+\]
+!et
+In case we use another activation function than the logistic one, we need to evaluate other derivatives. 
 
 
 !split 
@@ -1137,7 +1187,7 @@ We will be using the sigmoid function $\sigma(x)$:
   
 $$ f(x) = \sigma(x) = \frac{1}{1 + e^{-x}} ,$$
   
-which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011.
+which is inspired by probability theory (see logistic regression) and was most commonly used until about 2011. See the discussion below concerning other activation functions.
 
 !split  
 ===== Layers =====
@@ -1343,6 +1393,8 @@ This has two important benefits:
 o Introducing stochasticity decreases the chance that the algorithm becomes stuck in a local minima.  
 o It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.  
 
+The various optmization  methods, with codes and algorithms,  are discussed in our lectures on "Gradient descent approaches":"https://compphysics.github.io/MachineLearning/doc/pub/Splines/html/Splines-bs.html".
+
 !split  
 =====  Regularization =====
   
@@ -1376,9 +1428,9 @@ calculate the gradient efficently.
 ===== Matrix  multiplication =====
   
 To more efficently train our network these equations are implemented using matrix operations.  
-The error in the output layer is calculated simply as  
+The error in the output layer is calculated simply as, with $\hat{t}$ being our targets,  
   
-$$ \delta_L = \hat{y} - y = (n_{inputs}, n_{categories}) .$$  
+$$ \delta_L = \hat{t} - \hat{y} = (n_{inputs}, n_{categories}) .$$  
   
 The gradient for the output weights is calculated as  
   
@@ -1408,9 +1460,6 @@ $$ \nabla b_{h} = \sum_{i=1}^{n_{inputs}} \delta_h = (n_{hidden}) .$$
 
 !bc pycod
 # to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our model
 from sklearn.metrics import accuracy_score
 
 # one-hot in numpy
@@ -2189,90 +2238,133 @@ plt.show()
 !ec
 
 
+
+
 !split 
 ===== Which activation function should I use? =====
 
-Backpropagation algorithm works by going from the output layer to the
-input layer, propagating the error gradient on the way. Once the algorithm has computed the gradient of the
-cost function with regards to each parameter in the network, it uses these gradients to update each
-parameter with a Gradient Descent step.
+The Back propagation algorithm we derived above works by going from
+the output layer to the input layer, propagating the error gradient on
+the way. Once the algorithm has computed the gradient of the cost
+function with regards to each parameter in the network, it uses these
+gradients to update each parameter with a Gradient Descent (GD) step.
 
-Unfortunately, gradients often get smaller and smaller as the algorithm progresses down to the lower
-layers. As a result, the Gradient Descent update leaves the lower layer connection weights virtually
-unchanged, and training never converges to a good solution. This is called the vanishing gradients
-problem. In some cases, the opposite can happen: the gradients can grow bigger and bigger, so many
-layers get insanely large weight updates and the algorithm diverges. This is the exploding gradients
-problem, which is mostly encountered in recurrent neural networks. More generally,
-deep neural networks suffer from unstable gradients, different layers may learn at widely different speeds
+
+Unfortunately for us, the gradients often get smaller and smaller as the
+algorithm progresses down to the first hidden layers. As a result, the
+GD update leaves the lower layer connection weights
+virtually unchanged, and training never converges to a good
+solution. This is known in the literature as 
+_the vanishing gradients problem_. 
+
+In other cases, the opposite can happen, namely the the gradients can grow bigger and
+bigger. The result is that many of the layers get large updates of the 
+weights the
+algorithm diverges. This is the _exploding gradients problem_, which is
+mostly encountered in recurrent neural networks. More generally, deep
+neural networks suffer from unstable gradients, different layers may
+learn at widely different speeds
 
 !split 
 ===== Is the Logistic activation function (Sigmoid)  our choice? =====
 
-Although this unfortunate behavior has been empirically observed for quite a while (it was one of the
-reasons why deep neural networks were mostly abandoned for a long time), it is only around 2010 that
-significant progress was made in understanding it. 
+Although this unfortunate behavior has been empirically observed for
+quite a while (it was one of the reasons why deep neural networks were
+mostly abandoned for a long time), it is only around 2010 that
+significant progress was made in understanding it.
 
-A paper titled _Understanding the Difficulty of Training Deep Feedforward Neural Networks_ by Xavier Glorot and Yoshua Bengio1 found a few suspects,
-including the combination of the popular logistic sigmoid activation function and the weight initialization
-technique that was most popular at the time, namely random initialization using a normal distribution with
-a mean of 0 and a standard deviation of 1. In short, they showed that with this activation function and this
-initialization scheme, the variance of the outputs of each layer is much greater than the variance of its
-inputs. Going forward in the network, the variance keeps increasing after each layer until the activation
-function saturates at the top layers. This is actually made worse by the fact that the logistic function has a
-mean of 0.5, not 0 (the hyperbolic tangent function has a mean of 0 and behaves slightly better than the
-logistic function in deep networks).
+A paper titled "Understanding the Difficulty of Training Deep
+Feedforward Neural Networks by Xavier Glorot and Yoshua Bengio":"http://proceedings.mlr.press/v9/glorot10a.html" found that
+the problems with the popular logistic
+sigmoid activation function and the weight initialization technique
+that was most popular at the time, namely random initialization using
+a normal distribution with a mean of 0 and a standard deviation of
+1. 
+
+They showed that with this activation function and this
+initialization scheme, the variance of the outputs of each layer is
+much greater than the variance of its inputs. Going forward in the
+network, the variance keeps increasing after each layer until the
+activation function saturates at the top layers. This is actually made
+worse by the fact that the logistic function has a mean of 0.5, not 0
+(the hyperbolic tangent function has a mean of 0 and behaves slightly
+better than the logistic function in deep networks).
 
 
 !split
 ===== The derivative of the Logistic funtion =====
 
 Looking at the logistic activation function, when inputs become large
-(negative or positive), the function saturates at 0 or 1, with a derivative extremely close to 0. Thus when
-backpropagation kicks in, it has virtually no gradient to propagate back through the network, and what
-little gradient exists keeps getting diluted as backpropagation progresses down through the top layers, so
-there is really nothing left for the lower layers.
+(negative or positive), the function saturates at 0 or 1, with a
+derivative extremely close to 0. Thus when backpropagation kicks in,
+it has virtually no gradient to propagate back through the network,
+and what little gradient exists keeps getting diluted as
+backpropagation progresses down through the top layers, so there is
+really nothing left for the lower layers.
 
-In their paper, Glorot and Bengio propose a way to significantly alleviate this problem. We need the
-signal to flow properly in both directions: in the forward direction when making predictions, and in the
-reverse direction when backpropagating gradients. We don’t want the signal to die out, nor do we want it
-to explode and saturate. For the signal to flow properly, the authors argue that we need the variance of the
-outputs of each layer to be equal to the variance of its inputs, and we also need the gradients to have
-equal variance before and after flowing through a layer in the reverse direction (please check out the
-paper if you are interested in the mathematical details). 
+In their paper, Glorot and Bengio propose a way to significantly
+alleviate this problem. We need the signal to flow properly in both
+directions: in the forward direction when making predictions, and in
+the reverse direction when backpropagating gradients. We don’t want
+the signal to die out, nor do we want it to explode and saturate. For
+the signal to flow properly, the authors argue that we need the
+variance of the outputs of each layer to be equal to the variance of
+its inputs, and we also need the gradients to have equal variance
+before and after flowing through a layer in the reverse direction.
 
 
-One of the insights in the 2010 paper by Glorot and Bengio was that the vanishing/exploding gradients
-problems were in part due to a poor choice of activation function. Until then most people had assumed
-that if Nature had chosen to use roughly sigmoid activation functions in biological neurons, they
-must be an excellent choice. But it turns out that other activation functions behave much better in deep
-neural networks, in particular the ReLU activation function, mostly because it does not saturate for
-positive values (and also because it is quite fast to compute).
+
+One of the insights in the 2010 paper by Glorot and Bengio was that
+the vanishing/exploding gradients problems were in part due to a poor
+choice of activation function. Until then most people had assumed that
+if Nature had chosen to use roughly sigmoid activation functions in
+biological neurons, they must be an excellent choice. But it turns out
+that other activation functions behave much better in deep neural
+networks, in particular the ReLU activation function, mostly because
+it does not saturate for positive values (and also because it is quite
+fast to compute).
 
 
 !split
 ===== The RELU function family =====
 
 The ReLU activation function suffers from a problem known as the dying
-ReLUs: during training, some neurons effectively die, meaning they stop outputting anything other than 0.
+ReLUs: during training, some neurons effectively die, meaning they
+stop outputting anything other than 0.
+
+In some cases, you may find that half of your network’s neurons are
+dead, especially if you used a large learning rate. During training,
+if a neuron’s weights get updated such that the weighted sum of the
+neuron’s inputs is negative, it will start outputting 0. When this
+happen, the neuron is unlikely to come back to life since the gradient
+of the ReLU function is 0 when its input is negative.
+
+To solve this problem, nowadays practitioners use a  variant of the ReLU
+function, such as the leaky ReLU discussed above or the so-called
+exponential linear unit (ELU) function
 
-In some cases, you may find that half of your network’s neurons are dead, especially if you used a large
-learning rate. During training, if a neuron’s weights get updated such that the weighted sum of the neuron’s
-inputs is negative, it will start outputting 0. When this happen, the neuron is unlikely to come back to life
-since the gradient of the ReLU function is 0 when its input is negative.
 
-To solve this problem, you may want to use a variant of the ReLU function, such as the leaky ReLU discussed before or the so-called exponential linear unit (ELU) function
 !bt
 \[
 ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\  z & z \ge 0.\end{array}\right. 
 \]
 !et
 
-So which activation function should you use for the hidden layers of your deep neural networks? Although your mileage will vary,
-in general ELU is better than leaky ReLU (and its variants), which is better than ReLU. ReLU performs better than $\tanh$ which in turn performs better than the  logistic function. If you care a lot about runtime performance, then you
-may prefer leaky ReLUs over ELUs. If you don’t want to tweak yet another hyperparameter, you may just use the default $\alpha$ of
-$0.01$ for the leaky ReLU, and $1$ for ELU. If you have spare time and computing power, you can use
-cross-validation or bootstrap to evaluate other activation functions.
-huge training set.
+!split
+===== Which activation function should we use? =====
+
+In general it seems that the ELU activation function is better than
+the leaky ReLU function (and its variants), which is better than
+ReLU. ReLU performs better than $\tanh$ which in turn performs better
+than the logistic function. 
+
+If runtime
+performance is an issue, then you may opt for the  leaky ReLU function  over the 
+ELU function If you don’t
+want to tweak yet another hyperparameter, you may just use the default
+$\alpha$ of $0.01$ for the leaky ReLU, and $1$ for ELU. If you have
+spare time and computing power, you can use cross-validation or
+bootstrap to evaluate other activation functions.
 
 
 !split 
@@ -2297,8 +2389,9 @@ cardinal sin in ML. Then:
 * Make sure you are not overfitting.
 
 If the validation and test sets are drawn from the same distributions,
-then good performance on the validation set should lead to similarly
+then a good performance on the validation set should lead to similarly
 good performance on the test set. 
+
 However, sometimes
 the training data and test data differ in subtle ways because, for
 example, they are collected using slightly different methods, or