dpVAEs: Fixing Sample Generation for Regularized VAEs

TL;DR

This paper introduces decoupled priors (dpVAEs), a novel approach that separates representation learning from sample generation in regularized VAEs, improving their generative quality without losing representation benefits.

Contribution

The paper proposes a new family of priors, decoupled priors, that enable regularized VAEs to improve representation learning while maintaining high-quality sample generation.

Findings

dpVAE fixes sample generation issues in regularized VAEs.

Experiments on MNIST, SVHN, and CelebA show improved sample quality.

Decoupled priors can be integrated with existing regularizers without extra tuning.

Abstract

Unsupervised representation learning via generative modeling is a staple to many computer vision applications in the absence of labeled data. Variational Autoencoders (VAEs) are powerful generative models that learn representations useful for data generation. However, due to inherent challenges in the training objective, VAEs fail to learn useful representations amenable for downstream tasks. Regularization-based methods that attempt to improve the representation learning aspect of VAEs come at a price: poor sample generation. In this paper, we explore this representation-generation trade-off for regularized VAEs and introduce a new family of priors, namely decoupled priors, or dpVAEs, that decouple the representation space from the generation space. This decoupling enables the use of VAE regularizers on the representation space without impacting the distribution used for sample generation, and thereby reaping the representation learning benefits of the regularizations without sacrificing the sample generation. dpVAE leverages invertible networks to learn a bijective mapping from an arbitrarily complex representation distribution to a simple, tractable, generative distribution. Decoupled priors can be adapted to the state-of-the-art VAE regularizers without additional hyperparameter tuning. We showcase the use of dpVAEs with different regularizers. Experiments on MNIST, SVHN, and CelebA demonstrate, quantitatively and qualitatively, that dpVAE fixes sample generation for regularized VAEs.