Towards Deeper Understanding of Variational Autoencoding Models

TL;DR

This paper introduces a generalized optimization framework for variational autoencoders, providing theoretical insights into their behavior and proposing a new sequential VAE model that generates sharper images and learns more informative latent features.

Contribution

It generalizes the evidence lower bound, offers conditions for optimal data and feature learning, and introduces a new sequential VAE with improved image quality and latent feature learning.

Findings

The new criteria recover the data distribution under certain conditions.

The sequential VAE generates sharper LSUN images.

The proposed optimization encourages informative latent features.

Abstract

We propose a new family of optimization criteria for variational auto-encoding models, generalizing the standard evidence lower bound. We provide conditions under which they recover the data distribution and learn latent features, and formally show that common issues such as blurry samples and uninformative latent features arise when these conditions are not met. Based on these new insights, we propose a new sequential VAE model that can generate sharp samples on the LSUN image dataset based on pixel-wise reconstruction loss, and propose an optimization criterion that encourages unsupervised learning of informative latent features.