On Disentangled Representations Learned From Correlated Data

TL;DR

This paper investigates how popular disentanglement methods perform on correlated data, revealing that they often learn and reflect dataset correlations, and proposes solutions to mitigate this issue for better real-world applicability.

Contribution

It provides a large-scale empirical analysis of disentanglement methods on correlated data and offers strategies to reduce learned correlations using supervision or post-hoc corrections.

Findings

Disentanglement models learn and reflect dataset correlations.

Correlations in data affect downstream fairness applications.

Weak supervision and post-hoc correction can mitigate learned correlations.

Abstract

The focus of disentanglement approaches has been on identifying independent factors of variation in data. However, the causal variables underlying real-world observations are often not statistically independent. In this work, we bridge the gap to real-world scenarios by analyzing the behavior of the most prominent disentanglement approaches on correlated data in a large-scale empirical study (including 4260 models). We show and quantify that systematically induced correlations in the dataset are being learned and reflected in the latent representations, which has implications for downstream applications of disentanglement such as fairness. We also demonstrate how to resolve these latent correlations, either using weak supervision during training or by post-hoc correcting a pre-trained model with a small number of labels.