Refs

• https://deepgenerativemodels.github.io/notes/vae/ - excellent, concise
• https://ermongroup.github.io/cs228-notes/inference/variational/
• https://towardsdatascience.com/variational-inference-the-basics-f70ac511bcea
• https://stillbreeze.github.io/Variational-Inference-and-Expectation-Maximization/

In much of the following, I’m interchangeably using Θ and $z$.

Why latent variables

We want to model $p(X;\theta)$, but we also explicitly model latent variables/features $Z$.

Example: GMMs—for clustering, unsupervised learning

Another motivation:

Can also use more concrete examples if they’re easier to think about:

Overview

tldr:

• optimize parameters of a simpler distribution (from the variational family) to best approximate the target posterior distribution