Auto-Encoding Variational Bayes

Why this one

Read this after Maximum Likelihood from Incomplete Data via the EM Algorithm if you want the old probabilistic story, and before Generative Adversarial Networks or Denoising Diffusion Probabilistic Models if you want to see what later generators were reacting to. The key move is not “an encoder plus a decoder,” which is how VAEs are often taught. The key move is making a random sample differentiable by expressing it as deterministic parameters plus external noise. That reparameterization trick turns a messy latent variable model into something you can train with backprop. People also get the loss wrong by treating the KL term as decorative regularization. It is the price of forcing your learned code distribution to stay usable as a generative prior. Worth your evening because it teaches a pattern you will reuse everywhere: when gradients cannot pass through a thing, change the computation, not the goal.