Learning Disentangled Joint Continuous and Discrete Representations

TL;DR

This paper introduces an unsupervised framework for learning disentangled, interpretable joint continuous and discrete representations using variational autoencoders, effectively discovering underlying data factors.

Contribution

It proposes augmenting VAEs with relaxed discrete distributions and information control to automatically disentangle continuous and categorical factors.

Findings

Successfully disentangles continuous and discrete factors on multiple datasets.

Outperforms existing methods in scenarios with prominent discrete factors.

Demonstrates automatic discovery of generative factors without supervision.

Abstract

We present a framework for learning disentangled and interpretable jointly continuous and discrete representations in an unsupervised manner. By augmenting the continuous latent distribution of variational autoencoders with a relaxed discrete distribution and controlling the amount of information encoded in each latent unit, we show how continuous and categorical factors of variation can be discovered automatically from data. Experiments show that the framework disentangles continuous and discrete generative factors on various datasets and outperforms current disentangling methods when a discrete generative factor is prominent.