Disentangling by Factorising

TL;DR

This paper introduces FactorVAE, an unsupervised method for learning disentangled representations by promoting factorial distributions, improving upon beta-VAE, and proposing a new disentanglement metric.

Contribution

The paper presents FactorVAE, a novel approach that enhances disentanglement in representations and introduces a more reliable metric for evaluation.

Findings

FactorVAE achieves better disentanglement-reconstruction trade-off than beta-VAE.

A new disentanglement metric is proposed that overcomes limitations of previous metrics.

FactorVAE outperforms existing methods in disentangling independent factors.

Abstract

We define and address the problem of unsupervised learning of disentangled representations on data generated from independent factors of variation. We propose FactorVAE, a method that disentangles by encouraging the distribution of representations to be factorial and hence independent across the dimensions. We show that it improves upon $\beta$-VAE by providing a better trade-off between disentanglement and reconstruction quality. Moreover, we highlight the problems of a commonly used disentanglement metric and introduce a new metric that does not suffer from them.