Overlooked Implications of the Reconstruction Loss for VAE Disentanglement

TL;DR

This paper reveals that the interaction between data and the reconstruction loss in VAEs significantly influences disentanglement, challenging the common focus on regularisation and highlighting the subjective nature of disentanglement outcomes.

Contribution

The work uncovers the overlooked role of the reconstruction loss in VAE disentanglement and introduces a theory for adversarial datasets that hinder disentanglement.

Findings

Standard datasets have unintended correlations affecting disentanglement.

Constructed datasets can prevent disentanglement despite human-understandable factors.

A new reconstruction loss can restore disentanglement by aligning with ground-truth factors.

Abstract

Learning disentangled representations with variational autoencoders (VAEs) is often attributed to the regularisation component of the loss. In this work, we highlight the interaction between data and the reconstruction term of the loss as the main contributor to disentanglement in VAEs. We show that standard benchmark datasets have unintended correlations between their subjective ground-truth factors and perceived axes in the data according to typical VAE reconstruction losses. Our work exploits this relationship to provide a theory for what constitutes an adversarial dataset under a given reconstruction loss. We verify this by constructing an example dataset that prevents disentanglement in state-of-the-art frameworks while maintaining human-intuitive ground-truth factors. Finally, we re-enable disentanglement by designing an example reconstruction loss that is once again able to perceive the ground-truth factors. Our findings demonstrate the subjective nature of disentanglement and the importance of considering the interaction between the ground-truth factors, data and notably, the reconstruction loss, which is under-recognised in the literature.