Tutorial on Deep Learning III
summary
Compares undirected deep Boltzmann machines and directed variational autoencoders, explains why mean-field approximation gives fast, unimodal posterior estimates while MCMC/Gibbs samplers risk biased gradients if chains don’t mix, and highlights the reparameterization trick that lets VAEs backpropagate through stochastic layers. Demonstrations include generation of 1,500 characters from 50 alphabets, a conditional VAE trained on ~1M Flickr images (25k labeled helped classification), an IWAE bound that tightens as K increases, and failure modes like mode collapse and dataset priors (Obama/toilet-seat examples).
AI-written summary · about
chapters (3)
related talks
01h 22mDeep Probabilistic Methods with PyTorch
01h 17mDeep generative models for image...
28mPosterior Collapse in Deep Generative models
38mContinuous Representation Of Molecules using Graph Variational Autoencoder
39mBayesian Deep Learning with Edward (and a trick using Dropout)
35mVariational Inference and Python
37mVariational Inference in Python
41mBuilding Generative Image & Video models at Scale