Disentangling One Factor at a Time

Vaishnavi S Patil, Matthew S Evanusa, Joseph JaJa

Keywords: unsupervised representation learning, disentanglement, Variational Autoencoders, Generative Adversarial Networks

Abstract: With the overabundance of data for machines to process in the current state of machine learning, data discovery, organization, and interpretation of the data becomes a critical need. Specifically of need are unsupervised methods that do not require laborious labeling by human observers. One promising approach to this enedeavour is \textit{Disentanglement}, which aims at learning the underlying generative latent factors of the data. The factors should also be as human interpretable as possible for the purposes of data discovery. \textit{Unsupervised disentanglement} is a particularly difficult open subset of the problem, which asks the network to learn on its own the generative factors without any link to the true labels. This problem area is currently dominated by two approaches: Variational Autoencoder and Generative Adversarial Network approaches. While GANs have good performance, they suffer from difficulty in training and mode collapse, and while VAEs are stable to train, they do not perform as well as GANs in terms of interpretability. In current state of the art versions of these approaches, the networks require the user to specify the number of factors that we expect to find in the data. This limitation prevents "true" disentanglement, in the sense that learning how many factors is actually one of the tasks we wish the network to solve. In this work we propose a novel network for unsupervised disentanglement that combines the stable training of the VAE with the interpretability offered by GANs without the training instabilities. We aim to disentangle interpretable latent factors "one at a time", or OAT factor learning, making no prior assumptions about the number or distribution of factors, in a completely unsupervised manner. We demonstrate its quantitative and qualitative effectiveness by evaluating the latent representations learned on two benchmark datasets, DSprites and CelebA.

One-sentence Summary: Unsupervised disentanglement of one factor at a time from the entangled representations