Neural Discrete Representation Learning

TL;DR

This paper introduces VQ-VAE, a generative model that learns discrete representations to improve unsupervised learning and generation of high-quality images, videos, and speech, addressing issues like posterior collapse in VAEs.

Contribution

The paper presents VQ-VAE, a novel model that combines vector quantisation with variational autoencoders to learn discrete latent representations, enhancing generative capabilities.

Findings

Addresses posterior collapse in VAEs

Generates high-quality images, videos, and speech

Enables unsupervised phoneme learning

Abstract

Learning useful representations without supervision remains a key challenge in machine learning. In this paper, we propose a simple yet powerful generative model that learns such discrete representations. Our model, the Vector Quantised-Variational AutoEncoder (VQ-VAE), differs from VAEs in two key ways: the encoder network outputs discrete, rather than continuous, codes; and the prior is learnt rather than static. In order to learn a discrete latent representation, we incorporate ideas from vector quantisation (VQ). Using the VQ method allows the model to circumvent issues of "posterior collapse" -- where the latents are ignored when they are paired with a powerful autoregressive decoder -- typically observed in the VAE framework. Pairing these representations with an autoregressive prior, the model can generate high quality images, videos, and speech as well as doing high quality speaker conversion and unsupervised learning of phonemes, providing further evidence of the utility of the learnt representations.