Disentangling Domain and Content

TL;DR

This paper introduces a probabilistic model based on Variational Autoencoders that unsupervisedly disentangles domain and content features in data, enabling flexible image generation without explicit labels.

Contribution

It presents a novel, generalizable probabilistic framework for unsupervised disentanglement of domain and content, with state-of-the-art image generation performance.

Findings

Achieves high-quality image synthesis by combining domain and content from different inputs.

Operates effectively in few-shot, unsupervised settings without explicit labels.

Uses a novel domain-confusion loss to learn disentangled representations.

Abstract

Many real-world datasets can be divided into groups according to certain salient features (e.g. grouping images by subject, grouping text by font, etc.). Often, machine learning tasks require that these features be represented separately from those manifesting independently of the grouping. For example, image translation entails changing the style of an image while preserving its content. We formalize these two kinds of attributes as two complementary generative factors called "domain" and "content", and address the problem of disentangling them in a fully unsupervised way. To achieve this, we propose a principled, generalizable probabilistic model inspired by the Variational Autoencoder. Our model exhibits state-of-the-art performance on the composite task of generating images by combining the domain of one input with the content of another. Distinctively, it can perform this task in a few-shot, unsupervised manner, without being provided with explicit labelling for either domain or content. The disentangled representations are learned through the combination of a group-wise encoder and a novel domain-confusion loss.