Notes · [Date]

What a generative model actually learns

Draft. Outline only. Write the post under these headings.
A density, learned. A 10-component Gaussian mixture fitted to the two-moons dataset; contours show the learned log-density.

Outline

  1. A generative model is a density you can sample from
  2. Mixtures, autoregressive models, VAEs and flows on one toy dataset
  3. How to check a model: likelihood, samples and statistics
  4. Notes from working through Tomczak's book