Learning Hierarchical Priors in VAEs

TL;DR

This paper introduces a hierarchical prior in variational autoencoders to improve latent representations, using a graph-based interpolation method to align the latent space topology with the data manifold.

Contribution

It extends Taming VAEs to two-level hierarchical models and proposes a graph-based interpolation method for better latent space topology learning.

Findings

Latent representations reflect the data manifold topology

Hierarchical priors improve the interpretability of latent factors

The method learns smooth and explanatory latent features

Abstract

We propose to learn a hierarchical prior in the context of variational autoencoders to avoid the over-regularisation resulting from a standard normal prior distribution. To incentivise an informative latent representation of the data, we formulate the learning problem as a constrained optimisation problem by extending the Taming VAEs framework to two-level hierarchical models. We introduce a graph-based interpolation method, which shows that the topology of the learned latent representation corresponds to the topology of the data manifold---and present several examples, where desired properties of latent representation such as smoothness and simple explanatory factors are learned by the prior.