Auto-Encoding Variational Bayes
ReadDerives a tractable lower bound for training a latent-variable generative model by gradient descent. The probabilistic groundwork under diffusion models.
My notes
The hardest read of the first batch. I needed a pen for the ELBO derivation, and then the reparameterization trick made the whole thing suddenly feel obvious.
Shape of the Model
ELBO = E_q[ log p(x|z) ] − KL( q(z|x) ‖ p(z) )
In Plain Words
- First term: reconstruct the input well. Second term: keep the latent codes close to a simple prior so you can sample from it later.
- You cannot backprop through "sample z". You can backprop through "z = μ + σ·ε" — that one line is the paper's real contribution to practice.
- Blurry samples come from the Gaussian likelihood averaging over plausible outputs. GANs (next on the list) attack exactly that weakness.