DOT-VAE: Disentangling One Factor at a Time

TL;DR

DOT-VAE introduces a novel unsupervised framework for disentangling independent generative factors in real-world and synthetic datasets by augmenting variational autoencoders with a disentangled space and a two-step training process.

Contribution

It extends disentanglement methods to real-world data without prior assumptions, using a Wake-Sleep-inspired algorithm to learn interpretable factors one at a time.

Findings

Effective disentanglement on synthetic datasets DSprites and 3DShapes

Successful application to real-world CelebA dataset

No prior assumptions about number or distribution of factors

Abstract

As we enter the era of machine learning characterized by an overabundance of data, discovery, organization, and interpretation of the data in an unsupervised manner becomes a critical need. One promising approach to this endeavour is the problem of Disentanglement, which aims at learning the underlying generative latent factors, called the factors of variation, of the data and encoding them in disjoint latent representations. Recent advances have made efforts to solve this problem for synthetic datasets generated by a fixed set of independent factors of variation. Here, we propose to extend this to real-world datasets with a countable number of factors of variations. We propose a novel framework which augments the latent space of a Variational Autoencoders with a disentangled space and is trained using a Wake-Sleep-inspired two-step algorithm for unsupervised disentanglement. Our network learns to disentangle interpretable, independent factors from the data ``one at a time", and encode it in different dimensions of the disentangled latent space, while making no prior assumptions about the number of factors or their joint distribution. We demonstrate its quantitative and qualitative effectiveness by evaluating the latent representations learned on two synthetic benchmark datasets; DSprites and 3DShapes and on a real datasets CelebA.