Unifying Disentangled Representation Learning with Compositional Bias

Whie Jung, Dong Hoon Lee, Seunghoon Hong

Keywords: Unsupervised Representation Learning, Disentangled Representation Learning, Compositionality

Abstract:

Existing disentangled representation learning methods rely on inductive biases tailored for the specific factors of variation (e.g., attributes or objects).
However, these biases are incompatible with other classes of factors, limiting their applicability for disentangling general factors of variation.
In this paper, we propose a unified framework for disentangled representation learning, accommodating both attribute and object disentanglement.
To this end, we reformulate disentangled representation learning as maximizing the compositionality of the latents.
Specifically, we randomly mix two latent representations from distinct images and maximize the likelihood of the resulting composite image.
Under this general framework, we demonstrate that adjusting the strategy for mixing between two latent representations allows us to capture either attributes or objects within a single framework.
To derive appropriate mixing strategies, we analyze the compositional structures of both attributes and objects, then incorporate these structures into their respective mixing strategies.
Our evaluations show that our method surpasses or is comparable to state-of-the-art baselines such as DisDiff in attribute disentanglement (DCI, FactorVAE scores), and LSD and L2C in object property prediction tasks for object disentanglement.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6613