VMI-VAE: Variational Mutual Information Maximization Framework for VAE With Discrete and Continuous Priors

TL;DR

This paper introduces VMI-VAE, a framework that maximizes mutual information in variational autoencoders to improve the quality of learned representations and provides a way to evaluate this mutual information.

Contribution

It proposes a novel mutual information maximization framework for VAE that acts as a regularizer and enables mutual information evaluation.

Findings

Enhances the informativeness of latent codes in VAE.

Provides a method to evaluate mutual information in fixed VAE models.

Improves representation quality without sacrificing scalability.

Abstract

Variational Autoencoder is a scalable method for learning latent variable models of complex data. It employs a clear objective that can be easily optimized. However, it does not explicitly measure the quality of learned representations. We propose a Variational Mutual Information Maximization Framework for VAE to address this issue. It provides an objective that maximizes the mutual information between latent codes and observations. The objective acts as a regularizer that forces VAE to not ignore the latent code and allows one to select particular components of it to be most informative with respect to the observations. On top of that, the proposed framework provides a way to evaluate mutual information between latent codes and observations for a fixed VAE model.