Auto-Encoding Variational Bayes

Auto-Encoding Variational Bayes makes latent-variable models trainable with ordinary backpropagation. Read it to see how the evidence lower bound, amortized inference, and the reparameterization trick turn an intractable posterior problem into a practical stochastic-optimization recipe.

Reading focus: How the evidence lower bound replaces an intractable log-likelihood with reconstruction and KL terms that can be optimized. Why an encoder amortizes posterior inference by predicting a distribution over latent variables for each example. How reparameterizing a sample as a deterministic transform of parameter-free noise lets gradients pass through stochastic latent variables.

ICLR 2014. Kingma and Welling. 55 min read, hard difficulty.