@@ -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
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
@@ -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.
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
+$$
+\mathcal{C}(\hat{\beta}) = - \sum_{i=1}^n \left(y_i\log{p(y_i \vert x_i,\hat{\beta})}+(i-y_i)\log{1-p(y_i \vert x_i,\hat{\beta})}\right),
+$$
-
-
+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.
+
+
@@ -292,7 +328,7 @@ One can identify a set of key steps when using neural networks to solve supervis
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
-\( Y = (n_{inputs}) = (5, 3, 1, 8,...) \).
+
-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
-importnumpyasnp
-importmatplotlib.pyplotasplt
-fromsklearnimport 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 inenumerate(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()
-
-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.
-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.
+
+
-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
+deftrain_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
-
-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.
+
+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
-
-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
+
-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
-
-defsigmoid(x):
- return1/(1+ np.exp(-x))
-
-deffeed_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
-defpredict(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]))
-
-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
+
+defsigmoid(x):
+ return1/(1+ np.exp(-x))
+
+deffeed_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
+defpredict(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
-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:
-
-
-
Introducing stochasticity decreases the chance that the algorithm becomes stuck in a local minima.
-
It significantly speeds up the calculation, since we do not have to use the entire dataset to calculate the gradient.
-
+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:
-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:
+
+
+
Introducing stochasticity decreases the chance that the algorithm becomes stuck in a local minima.
+
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.
@@ -309,7 +320,7 @@ calculate the gradient efficently.
-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
-fromsklearn.metricsimport accuracy_score
-
-# one-hot in numpy
-defto_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)
-
-deffeed_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
-
-defbackpropagation(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 inrange(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)))
-
-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
+fromsklearn.metricsimport accuracy_score
+
+# one-hot in numpy
+defto_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)
+
+deffeed_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
+
+defbackpropagation(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 inrange(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
-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 \( \lambda = 10^{-6},...,10^{-0} \).
-
-
+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
+
-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.
-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)
+
+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
-importseabornassns
+
-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
-
fromsklearn.neural_networkimport MLPClassifier
-# store models for later use
-DNN_scikit = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
+
+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
-importseabornassns
+
-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.
+
+
-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
+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
-importnumpyasnp
-importmatplotlib.pyplotasplt
-fromsklearnimport 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 inenumerate(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)
+
-
fromkeras.utilsimport to_categorical
-fromsklearn.model_selectionimport 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)
+
-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:
+
# optional
+# we can use log files to visualize our graph in Tensorboard
+writer = tf.summary.FileWriter('logs/')
+writer.add_graph(tf.get_default_graph())
+
-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
+
+
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).
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).
-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).
-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
+
+$$
+ELU(z) = \left\{\begin{array}{cc} \alpha\left( \exp{(z)}-1\right) & z < 0,\\ z & z \ge 0.\end{array}\right.
+$$
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
+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.
+
+
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.
+
+
-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
-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
@@ -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
+
+
+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
+
+
+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
+
+
+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,
# to categorical turns our integer vector into a onehot representation
-#from keras.utils import to_categorical
-
-# calculate the accuracy score of our modelfromsklearn.metricsimport 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
+
-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 modelfromsklearn.metricsimport 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
+
-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 modelfromsklearn.metricsimport 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
+
-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": "iVBORw0KGgoAAAANSUhEUgAAAYgAAAEXCAYAAAC3c9OwAAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDMuMC4wLCBodHRwOi8vbWF0cGxvdGxpYi5vcmcvqOYd8AAAGs5JREFUeJzt3Xu0HWWZ5/HvOSFcEyIeDk0AI9ojj9q0BCHY0wjaQ8aepp1WR4SWTEdRQBRtL83ysrg02IOt9nKiLi66UMQW4wVUehSQHgYUEGmaUVAbfGhbQCdkNB6yhgS5JDl7/qg6nO2xQk52Vaid4vtZK4tdVW/tevJycn77rbeq9kiv10OSpJlG2y5AkjScDAhJUiUDQpJUyYCQJFUyICRJlQwISVIlA0KdFxH/GBF7Nvh+p0fEzyLiM029Z/m+CyLiur7l2yPiaU0eQ9oaO7RdgPQk+I8Nv98bgeMz86aG33cP4LCphcxc3PD7S1tlxBvl1AURMQ/4DPAcYBL438CbgE8Drwd+BBxdbjsPWATMBb6YmR+IiP2BbwPfAg4CRoC3ZuaNM47zJeBVwD3AWcCbgfMy8/Jy+7emliPiEeCDFAG1D/CxzPxo2e59wOuAjcC/ljVeARwJ/BA4pNw2npm/iogzgdeW6+4ua/u/5fG+Cxxe/p1uBF6XmZP1elTyFJO641XA/PJT95Jy3bMz84Ty9R9l5s+BzwEXZ+YhFJ/Wl0bEsWWbRcA15Xu8F/hSRMztP0hmHgfcDyzLzC9toaadgF9l5uHAMcAHI2LniPgzikD495l5IEXYvBU4AXg4Mxdn5qapN4mIE4A/AZZk5gsowu6SvuP8LvBS4PeB/wC8ZAt1SbNiQKgrbgJ+r/xE/V7go5n5k/4GEbEbxS/Pv4mI24FbKEJh6lTO2sxcCZCZVwObgBfUrOsfyv9+jyIwdgOWApdl5tryWO/KzHOf4D3+BPhMZj5ULn8MOCoidiyXv56Zk5m5DvgJ8PSaNUuAAaGOyMx7gH8H/C2wO3BtRBwzo9kcilNHf1h+Sl8M/AHwgXL7xhntRylC4on0yvecsuOM7Q+X9U2dyx0pj/P4ud2IeFp5imtzZv47HaWYP5w67sNPUI80MANCnRARb6aYg/jHzHwPcA1wYLl5EzA3Mx+kGDW8q9znacB3gFeU7cYj4j+V2/4zsIFiPuCJrAEOLff5XWY34rgW+C8RsXu5fHZZ00ZgTkTM/AV/DXBCOQIC+Evghsx8dBbHkgZmQKgr/p5ihHBnRNxGMYr4WLntq8BNEXEgcDzwBxHxQ+CfgC9k5ufLdo8AfxERdwCnA6/snwvYjP8GvCwifgR8CLhhS4Vm5lUUYfadso69y+OtpjgVdVdEjPXt8mmKULk1Iu4CXggs29JxpLq8ikkCylM8P8rMeW3XIg0LRxCSpEq1RxDledSbgZdn5r0ztp1FcVPR2nLVRZl5fkQsBi4CFlAMyU/JzJkThJKkFtW6kzoiXkTxi/6AzTRZAvx5Zn53xvpLgRMz85aI+DRwEnBhnVokSc2q+6iNk4BTKW4+qnIo8J6IeDbFSOE04HeAXTLzlrLNJcA5GBCSNFRqBURmnggQEb+1rXz0wfcpQuFeiiA4E/gGxdUaU1YD+83ykDtRjEpWs+Xr0yVJhTnAQuCfgVlfHr3NHtaXmespnn0DQER8BLgYuLKi+WyfG7OE4lkzkqStdwTFUwdmZZsFREQsApZm5sXlqhGKG49WUVz3PWUhxbNtZmM1wNq1DzE52e7luWNj85iYWN9qDcPCvphmX0yzL6a13RejoyPsscdu8Jtnb7ZoWz7u+2HgwxFxPcUpplOBr2XmfRHxSEQcnpnfAZYDV8/yPTcBTE72Wg+IqTpUsC+m2RfT7ItpQ9IXW3VqvvH7ICLiqog4NDPXUDxu+etAUowgPlI2WwasKO8K3Q34eNN1SJLq2d7upN4fuGdiYn3raTw+Pp81a9a1WsOwsC+m2RfT7ItpbffF6OgIY2PzAJ5FcUZndvttq4IkSds3A0KSVMmAkCRVMiAkSZUMCElSJQNCklTJgJAkVTIgJEmVDAhJUiUDQpJUyYCQJFUyICRJlQwISVIlA0KSVMmAkCRVMiAkSZUMCElSJQNCklRph7pvEBG7AzcDL8/Me2dsewVwDsX3Ud8DnJCZayNiOfAh4Bdl0ysz8/S6tUiSmlMrICLiRcBFwAEV23YHLgSWZOaqiHg/cDbwdmAJ8K7M/EKd40uStp26p5hOAk4F7q/YNhd4S2auKpd/ACwqXy8BlkfEHRFxaUTsUbMOSVLDagVEZp6YmTduZttEZl4BEBG7AO8Frig3r6YYTSwGfg6cV6cOSVLzRnq9Xu03iYh7gZfOnIMoty2gCIafZuYbK7bvUW6bzShif4q5DEnS1nsWcO9sG9eepH4iEbEQuAa4DnhnuW4B8IbMXFE2GwE2bM37TkysZ3KyfrDVMT4+nzVr1rVaw7CwL6bZF9Psi2lt98Xo6AhjY/O2fr9tUAsAETEH+Abw5cx8R2ZO/UZfD7y7nOAGeCvwtW1VhyRpMI2PICLiKuAs4BnAwcCciDim3HxbZp4YEccCF5ZzE3cDy5uuQ5JUTyMBkZn7970+unx5G5sZoZQT2y9s4tiSpG3DO6klSZUMCElSJQNCklTJgJAkVTIgJEmVDAhJUiUDQpJUyYCQJFUyICRJlQwISVIlA0KSVMmAkCRVMiAkSZUMCElSJQNCklTJgJAkVTIgJEmVDAhJUiUDQpJUqfZ3UkfE7sDNwMsz894Z2xYDFwELgBuAUzJzY0QsAi4F9gISWJaZ6+vWIklqTq0RRES8CLgJOGAzTS4F3paZBwAjwEnl+guACzLzucBtwJl16pAkNa/uCOIk4FTgczM3RMQzgV0y85Zy1SXAORHxKeBI4JV9678NvKdmLVKrNm6a5MvX3s2v1j7UdilDYdddduTXDz/WdhlDoe2+2HPBzhz7sudt9X61AiIzTwSIiKrN+wCr+5ZXA/sBewIPZubGGetnbWxs3lbXui2Mj89vu4ShYV/AXfc8wOeuvos5oyOMjo60XY70uP32mv/kB8QWVP0LmXyC9bM2MbGeycneQEU1ZXx8PmvWrGu1hmFhXxQmHiim0d557EE8f/+nt1xN+/y5mNZ2Xwz6gWVbXsW0Cti7b3khcD+wBtg9IubMWC91wsiIowd1wzYLiMy8D3gkIg4vVy0Hrs7MDcCNwHH967dVHdKTZWpQazyoKxoPiIi4KiIOLReXASsi4i5gN+Dj5fq3ACdHxJ3AEcAZTdchPel6RUI4gFBXNDIHkZn7970+uu/1HcBhFe3vA17axLGlYTE1keYpJnWFd1JLDek5glDHGBBSU6bmIEwIdYQBITXESWp1jQEhNWbqFJMRoW4wIKSGPD6CMB/UEQaE1BQDQh1jQEgNefwqJmch1BEGhNQQTzGpawwIqTFOUqtbDAipIT0vc1XHGBBSQ6YeQO8AQl1hQEgN6fWchFC3GBBSQybLgPDL5NQVBoTUlHa/5FBqnAEhNWTqDNOop5jUEQaE1JAePu5b3WJASA3peRmTOsaAkBry+CmmdsuQGlPrK0cj4niK75PeEViRmef3bVsMXNLXfBxYm5kHRsRy4EPAL8ptV2bm6XVqkdrW8045dczAARER+wLnAocAjwI3R8T1mXknQGbeDiwu2+4K3AqcUu6+BHhXZn6hRu3SUJk6w+Qktbqizmh4KXBdZj6QmQ8BlwPHbKbt+4BvZ+ZN5fISYHlE3BERl0bEHjXqkIbC4yMIqSPqnGLaB1jdt7waOGxmo4h4GnAy8Psz2n6QYlTxAeA8YNlsDzw2Nm+Acps3Pj6/7RKGhn0Bu83bGYA995zH2IJdWq5mOPhzMW177Is6AVE1jp6sWLcMuCIzfzm1IjNfNfU6Ij4M/HRrDjwxsZ7JyXY/rY2Pz2fNmnWt1jAs7IvCunWPALD2gYeYfGxjy9W0z5+LaW33xejoyEAfrOucYloF7N23vBC4v6LdK4EvTi1ExIKIeGff9hFgQ406pKHgs5jUNXUC4lrgqIgYLyehXw18s79BRIxQTGJ/t2/1euDdEfGicvmtwNdq1CENBfNBXTNwQGTmKuB04HrgdmBlZt4aEVdFxKFls3Hgscx8pG+/TcCxwIURcRdFgLx70DqkYTH9laNSN9S6DyIzVwIrZ6w7uu/1L/nN01BT628EXljn2NKwmR5BGBHqBm/6lBoyfR9Eq2VIjTEgpIZM3wdhQqgbDAipIU5Sq2sMCKkhPu5bXWNASE1xklodY0BIDZn0Mld1jAEhNcTLXNU1BoTUEL9QTl1jQEgNefxOagNCHWFASE3xFJM6xoCQGuIktbrGgJAa5ghCXWFASA2Z7Dn/oG4xIKSG9Ho9Rw/qFANCapBPclWXGBBSQ4pJahNC3WFASE3pOYJQtxgQUkOKAYQJoe6o9ZWjEXE8cAawI7AiM8+fsf0s4I3A2nLVRZl5fkQsBi4CFgA3AKdk5sY6tUht69FzBKFOGXgEERH7AucCLwYOAk6OiOfPaLYE+PPMXFz+mQqQS4G3ZeYBFCdtTxq0DmlY9LzMVR1TZwSxFLguMx8AiIjLgWOA9/e1ORR4T0Q8m2KkcBrwO8AumXlL2eYS4Bzgwhq1SK2b9DJXdUydOYh9gNV9y6uB/aYWImIe8H2KUHgh8DTgzC3tJ223et5FrW6pM4Ko+pcwOfUiM9cDR08tR8RHgIuBK59ov9kYG5u3Nc23mfHx+W2XMDTsC9h557mMYF/0sy+mbY99UScgVgFH9C0vBO6fWoiIRcDSzLy4XDUCbCj323tz+83GxMR6Jid7W264DY2Pz2fNmnWt1jAs7IvCrx9+jJGREfui5M/FtLb7YnR0ZKAP1nVOMV0LHBUR4xGxK/Bq4Jt92x8GPhwRz4qIEeBU4GuZeR/wSEQcXrZbDlxdow5pKDhJra4ZOCAycxVwOnA9cDuwMjNvjYirIuLQzFwDvAn4OpAUI4iPlLsvA1ZExF3AbsDHa/wdpKHQA0ZNCHVIrfsgMnMlsHLGuqP7Xn8F+ErFfncAh9U5tjRser2eT9pQp3gntdSQXs8b5dQtBoTUkJ6XuapjDAipIZ5hUtcYEFJDevQY8RyTOsSAkBriCEJdY0BIDXEOQl1jQEgN6dHzRjl1igEhNcQRhLrGgJAa4n0Q6hoDQmpIrwdOU6tLDAipIT1g1H9R6hB/nKWG9Ho9HEGoSwwIqSG9Hs5BqFMMCKkhPb+TWh1jQEgN8QuD1DUGhNQgRxDqEgNCashkzzup1S0GhNQgRxDqklpfORoRxwNnADsCKzLz/BnbXwGcQ3Ht3z3ACZm5NiKWAx8CflE2vTIzT69Ti9S2yV7Pi1zVKQMHRETsC5wLHAI8CtwcEddn5p3l9t2BC4ElmbkqIt4PnA28HVgCvCszv1CzfmloFJe5GhHqjjqnmJYC12XmA5n5EHA5cEzf9rnAWzJzVbn8A2BR+XoJsDwi7oiISyNijxp1SMPBOQh1TJ2A2AdY3be8GthvaiEzJzLzCoCI2AV4L3BFX9uzgcXAz4HzatQhDYVJn+aqjqkzB1H1L2Fy5oqIWEARDHdk5mcBMvNVfds/DPx0aw48NjZv6yrdRsbH57ddwtCwL2Du3Dls3DRpX/SxL6Ztj31RJyBWAUf0LS8E7u9vEBELgWuA64B3lusWAG/IzBVlsxFgw9YceGJiPZOTvQHLbsb4+HzWrFnXag3Dwr4oPPbYRubsMMe+KPlzMa3tvhgdHRnog3WdU0zXAkdFxHhE7Aq8Gvjm1MaImAN8A/hyZr4jM6d+o68H3h0RLyqX3wp8rUYd0lDwTmp1zcAjiPLKpNOB6ykuc/1UZt4aEVcBZwHPAA4G5kTE1OT1bZl5YkQcC1xYzk3cDSyv9beQhkCv12PEC13VIbXug8jMlcDKGeuOLl/exmZGKJl5I/DCOseWhs0kjiDULd5JLTXF+yDUMQaE1JBer+f3BalTDAipIT0cQahbDAipIY4g1DUGhNQQn8WkrjEgpIb02r13U2qcASE1pNfrOYJQpxgQUkN6eB+EusWAkBrS83Hf6hgDQmpIMYIwIdQdBoTUEB/Wp64xIKSGFKeYTAh1hwEhNcT75NQ1BoTUEC9zVdcYEFJDnKRW1xgQUkO8zFVdY0BIDfEqJnWNASE1pAgIE0LdUesrRyPieOAMiu+kXpGZ58/Yvhi4CFgA3ACckpkbI2IRcCmwF5DAssxcX6cWqX2eYlK3DDyCiIh9gXOBFwMHASdHxPNnNLsUeFtmHkBxBeBJ5foLgAsy87kU31195qB1SMNisgcjXuiqDqlzimkpcF1mPpCZDwGXA8dMbYyIZwK7ZOYt5apLgNdExFzgyLL94+tr1CENBSep1TV1TjHtA6zuW14NHLaF7fsBewIPZubGGetn7aNfvoMHHnxkqwtu0g47zGHjxk2t1jAs7IvCul9vcA5CnVInIKr+JUzOYvuW9tuivcZ2Zaedak2fSI1bOD6PIw/el/Hx+W2XMjTsi2nbY1/U+S27Cjiib3khcP+M7XtXbF8D7B4RczJzU8V+W3T8Uc9hcrLdr+8aH5/PmjXrWq1hWNgX0+yLafbFtLb7YnR0hLGxeVu/X41jXgscFRHjEbEr8Grgm1MbM/M+4JGIOLxctRy4OjM3ADcCx/Wvr1GHJGkbGDggMnMVcDpwPXA7sDIzb42IqyLi0LLZMmBFRNwF7AZ8vFz/Foqrnu6kGIWcMWgdkqRtY6S3fX3T+v7APRMT6z3FNETsi2n2xTT7YlrbfdF3iulZwL2z3m9bFSRJ2r4ZEJKkSgaEJKmSASFJqmRASJIqGRCSpEoGhCSpkgEhSapkQEiSKhkQkqRKBoQkqZIBIUmqZEBIkioZEJKkSgaEJKmSASFJqmRASJIqGRCSpEoGhCSp0g6D7hgRi4BLgb2ABJZl5voZbRYCnwH2BiaB0zLzuoiYC0wAP+1rfkhmbhq0HklSs+qMIC4ALsjM5wK3AWdWtPk74BuZuRh4LbAyIuYALwC+m5mL+/4YDpI0RAYKiHIEcCRwebnqEuA1FU2/Cny+fP0TYGdgHrAEGI+IW8o/LxmkDknStjPoKaY9gQczc2O5vBrYb2ajzPxq3+JpwPcz8/9FRA+4AvgbYDFwdUQcmJm/ms3Bx8bmDVh2s8bH57ddwtCwL6bZF9Psi2nbY19sMSAi4jXAihmr765oOvkE7/EO4E3ASwAy85N9m78fEf8EHA78w5bqAZiYWM/kZG82TbeZ8fH5rFmzrtUahoV9Mc2+mGZfTGu7L0ZHRwb6YL3FgMjMy4DL+tdNTTJHxJxy7mAhcH/V/hHxYeBPgSMz8/+U6/4CuDkz/61sNgJs2OrqJUnbzEBzEJm5AbgROK5ctRy4ema7cuTwR8DhU+FQOgj4q7JNAAeX7ydJGhIDX+YKvAX4bEScAfyM4iolIuIUYB/gr8s/DwLfKnIAgKOB9wMXR8SPgB6wPDMdi0rSEBk4IDLzPuClFes/0be4xxO8xTGDHluStO15J7UkqZIBIUmqZEBIkioZEJKkSgaEJKmSASFJqmRASJIqGRCSpEoGhCSpkgEhSapkQEiSKhkQkqRKBoQkqZIBIUmqZEBIkioZEJKkSgaEJKmSASFJqjTwV45GxCLgUmAvIIFlmbm+os2/AP9WrvpFZv5xROwIfBo4FHgYOD4zfzxoLZKk5tUZQVwAXJCZzwVuA86saLMEWJmZi8s/f1yu/0vgocx8HvAO4LM16pAkbQMDjSAiYi5wJPDKctUlwLeB98xougQ4MCJuAx4E3p6ZPwT+FDgLIDNviIg9I2JRZv5sC4eeAzA6OjJI2Y0bljqGgX0xzb6YZl9Ma7Mv+o49Z2v2G/QU057Ag5m5sVxeDexX0e4R4O8z85MRcTRwRUQ8D9in3IcZ+28pIBYC7LHHbgOW3ayxsXltlzA07Itp9sU0+2LakPTFQqZP+W/RFgMiIl4DrJix+u6KppMzV2Tm2X2vr4qIvwWeB1RF6W/tX+GfgSMoAmXTLNpLkoqRw0KK36GztsWAyMzLgMv615WnmCYiYk5mbioPfP/MfSPibRRzEBPlqhFgA7AK2Bv4Sbm+cv8KjwI3zaKdJOk3zXrkMGWgSerM3ADcCBxXrloOXF3R9CXAGwEi4iUUKfZj4KpyHyLixcAjs5h/kCQ9iUZ6vd5AO0bEMymuPtqLYu7gtZm5NiJOAfbJzLMiYl+KCeyFFJezvjEzfxAROwOfpLjM9VHgxMz8Xu2/jSSpMQMHhCSp27yTWpJUyYCQJFUyICRJlQwISVKlgR/Wp0JEHAzckpk7tV1LWyLicOCjwFxgAnhDZt7XblVProg4HjgD2BFYkZnnt1xSayLir4Fjy8UrM/PdbdYzDCLi74DxzHx927VsDUcQNUTErsB5FL8Unso+T3EJ8+Ly9cdbrudJVV7OfS7wYuAg4OSIeH67VbUjIpYCLwMOBhYDh0TEq9qtql0RcRTw+rbrGIQBUc9H+O3HkDylRMROwBmZ+YNy1Q+ARS2W1IalwHWZ+UBmPgRcDhzTck1tWQ38VWY+Vt5QexdPvZ+Hx0XE0yk+PHyg7VoG4SmmAUXEnwG7ZublEdF2Oa3JzEcpvheEiBgFzgauaLOmFlQ9fPKwlmppVWb+y9TriHgOxdMW/rC9ilr3SeB04BltFzIIA2ILNvOwwh8Du1N8cnzK2FxfZObS8kugPkvxM7VdflqqYdCHT3ZWRPwecCVwWmb+a9v1tCEiTgR+npn/KyJe33Y9g/BO6gGU/+PfB6wrVx0E3AEckZnrNrtjR0XEPOB/UExQ/9dyVPGUERGvo/h/f2K5fCYwkpnvb7eydpQXLXwFeEdmfrHtetoSEf+T4jFDG4GnA/OAz2bmO1stbCsYEA2IiF5mPmW/GSUirgB+CbwpM59yP1DlJPVNFKeVHgJuBk7OzFtbLawFEfEM4HvAcZl5Xdv1DItyBPHS7e0qJk8xqZbyMt9XAHcC3y/nY+7PzKNbLexJlJmrIuJ04HqKK9o+9VQMh9JpwM7Af++bm/tEZn6ivZI0KEcQkqRKXuYqSapkQEiSKhkQkqRKBoQkqZIBIUmqZEBIkioZEJKkSt4oJ9UUEacCJ/Wtej7wocw8s6WSpEZ4o5zUoIh4M/AGiscqPNR2PVIdjiCkhpRfjHMacLjhoC5wBCE1oHyC6WXA0sy8s+16pCY4SS3VVH696GXA8YaDusQRhFRTRFwDHALcy/Rp29umvh9C2l4ZEJKkSp5ikiRVMiAkSZUMCElSJQNCklTJgJAkVTIgJEmVDAhJUiUDQpJU6f8Dl0cLEFBoNX0AAAAASUVORK5CYII=\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": "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\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": "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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"
- }
- ],
+ "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 e55a4b7de..8fa98fbb3 100644
Binary files a/doc/pub/NeuralNet/ipynb/ipynb-NeuralNet-src.tar.gz and b/doc/pub/NeuralNet/ipynb/ipynb-NeuralNet-src.tar.gz differ
diff --git a/doc/pub/NeuralNet/pdf/NeuralNet-minted.pdf b/doc/pub/NeuralNet/pdf/NeuralNet-minted.pdf
index 96c52cb51..8ba639a9f 100644
Binary files a/doc/pub/NeuralNet/pdf/NeuralNet-minted.pdf and b/doc/pub/NeuralNet/pdf/NeuralNet-minted.pdf differ
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