Unsupervised Disentanglement without Autoencoding: Pitfalls and Future Directions

TL;DR

This paper investigates the challenges of unsupervised disentangled representation learning without autoencoders, highlighting limitations of current regularization methods and proposing future research directions.

Contribution

It critically analyzes the pitfalls of existing contrastive regularization approaches for disentanglement and discusses potential future strategies beyond autoencoding methods.

Findings

Unsupervised disentanglement is sensitive to optimization and initialization.

Contrastive regularization methods face trade-offs in disentanglement and task performance.

Autoencoder-based methods do not scale well to large datasets.

Abstract

Disentangled visual representations have largely been studied with generative models such as Variational AutoEncoders (VAEs). While prior work has focused on generative methods for disentangled representation learning, these approaches do not scale to large datasets due to current limitations of generative models. Instead, we explore regularization methods with contrastive learning, which could result in disentangled representations that are powerful enough for large scale datasets and downstream applications. However, we find that unsupervised disentanglement is difficult to achieve due to optimization and initialization sensitivity, with trade-offs in task performance. We evaluate disentanglement with downstream tasks, analyze the benefits and disadvantages of each regularization used, and discuss future directions.