Leveraging Hidden Structure in Self-Supervised Learning

TL;DR

This paper introduces a mutual information-based framework for learning structured, interpretable representations from raw images using self-supervised learning, demonstrating improved generalization and interpretability on CIFAR-10.

Contribution

It presents a novel framework combining self-supervised and structure learning with a post-hoc interpretability method, advancing structured representation learning.

Findings

Higher generalization performance on CIFAR-10 classification

More interpretable representations compared to traditional self-supervised methods

Framework effectively integrates structure learning with self-supervised learning

Abstract

This work considers the problem of learning structured representations from raw images using self-supervised learning. We propose a principled framework based on a mutual information objective, which integrates self-supervised and structure learning. Furthermore, we devise a post-hoc procedure to interpret the meaning of the learnt representations. Preliminary experiments on CIFAR-10 show that the proposed framework achieves higher generalization performance in downstream classification tasks and provides more interpretable representations compared to the ones learnt through traditional self-supervised learning.