DVAE++: Discrete Variational Autoencoders with Overlapping Transformations

TL;DR

DVAE++ introduces overlapping transformations for discrete latent variables, enabling effective training with Boltzmann priors and outperforming existing relaxations like Gumbel-Softmax in benchmark tests.

Contribution

The paper proposes a novel smoothing transformation for discrete variables and a variational bound, facilitating training of discrete VAEs with complex priors.

Findings

Overlapping transformations outperform Gumbel-Softmax in benchmarks.

The method effectively trains models with Boltzmann priors.

DVAE++ achieves superior generative performance on several datasets.

Abstract

Training of discrete latent variable models remains challenging because passing gradient information through discrete units is difficult. We propose a new class of smoothing transformations based on a mixture of two overlapping distributions, and show that the proposed transformation can be used for training binary latent models with either directed or undirected priors. We derive a new variational bound to efficiently train with Boltzmann machine priors. Using this bound, we develop DVAE++, a generative model with a global discrete prior and a hierarchy of convolutional continuous variables. Experiments on several benchmarks show that overlapping transformations outperform other recent continuous relaxations of discrete latent variables including Gumbel-Softmax (Maddison et al., 2016; Jang et al., 2016), and discrete variational autoencoders (Rolfe 2016).