https://xkcd.com/1399/
1
Linear and Nonlinear Component Analysis
Probabilistic PCA
Autoencoders
Autoencoders for image processing
Recommender systems
2
PCA: recap
3
min
Maximum variance projection
Minimum error formulation
Covariance matrix of
(centered) data
Probabilistic PCA
4
x : D-dimensional vector
z : M-dimensional Gaussian latent variable
D >= M
W: D-by-M matrix
Likelihood and posterior for x and z
5
Woodbury / matrix inversion identity
also (2.113)-(2.117)
D-by-D
M-by-M
