Reading Assignment 7: Normalizing Flows

Overview

The below are multiple options for getting the general idea of normalizing flows, along with an overview over different methods.

• An approachable introduction to normalizing flows can be found in this blog post by Lilian Weng. You can skip the parts on autoregressive models (PixelRNN, Wavenet).
• Next, this two-part blog post by Eric Jang provides another view; especially the discussion on the meaning of the Jacobian determinant in the first part can be a nice addition. A lot of information will be a repetition from the first blog. You can/should skip the code sections; they are heavily outdated.
• Finally, Murphy’s book has an in-depth chapter devoted to normalizing flows.
• If you need a refresher, on the substitution rule for integrals, the Wikipedia article should be sufficient. The change of variable theorem used for normalizing flows is the multi-dimensional version of this.

Specific Models

This can be considered optional reading, but try to read at least one of these to get a detailed view on a specific flow model.

• NICE: A relatively simple model and not very powerful, but it arguably lead the groundwork for all deep flow models.
• RealNVP: A massive step up from NICE.
• Glow: Yet another iteration on RealNVP, using invertible 1x1 convolutions.
• Parallel Wavenet: A neat application of IAFs for efficient Wavenet sampling.
• A whole other usage of normalizing flows is to improve the variational posteriors in VAEs.