Morten Hjorth-Jensen
35 lines
1.4 KiB
Plaintext
35 lines
1.4 KiB
Plaintext
Topics to consider adding:
|
||
|
||
a) add about kernel methods after having discussed ridge and lasso regression
|
||
|
||
1) words on Bayesian ML
|
||
|
||
A group of machine learning algorithms where the fits and
|
||
optimizations are based on Bayesian statistics (Bayes’ Theorem)
|
||
instead of traditional statistics and optimizations Assume outputs can
|
||
be described as distributions instead of as linear and nonlinear
|
||
combinations of inputs and hyperparameters are fit using priors
|
||
instead of being set by user Benefits: No hyperparameter tuning, no
|
||
validation data set, produced uncertainties on predictions
|
||
|
||
|
||
2) Discuss Kernel regression first and then add Bayesian Ridge regression
|
||
Bayesian Ridge Regression
|
||
Bayesian version of ridge regression (regularized linear regression)
|
||
Finds parameters and hyperparameters using Gaussian distributions
|
||
Different results than ridge regression but does not depend on user-ser hyperparameter
|
||
|
||
3) GP
|
||
Gaussian Processes
|
||
Bayesian version of kernel ridge regression or support vector machines
|
||
Similar to Bayesian ridge regression but uses the kernel trick to modify the inputs
|
||
Kernel: Modified Rational Quadratic Kernel
|
||
|
||
5) Update lectures on CNNs and RNNs
|
||
|
||
4) Add about generative models, AE, VAE and GANS, discuss energy based models and how to compute the partition function.
|
||
5) add discussions about Kullback-Leibner
|
||
|
||
|
||
Notes to self: look up thesis by Jane for RBMs and FNNs applied to many-body problems and more
|