Auto-Encoding Variational Bayes
The variational autoencoder trains an encoder to approximate the posterior over latent variables and a decoder to reconstruct data, maximising a lower bound on the likelihood. The reparameterisation trick makes the sampling step differentiable so the whole thing trains with ordinary gradients.
Key ideas
- Maximise a tractable lower bound instead of the intractable likelihood
- Sample by transforming noise so gradients pass through
- The KL term is a regulariser toward the prior
Why read it now
The evidence lower bound and the reparameterisation trick are the two tools that connect probability to backpropagation; diffusion models are their descendants.
Question to keep in mind
Which term of the objective pulls the latent space toward the prior, and what happens to samples if it wins?