Causes and Consequences of Representational Similarity in Machine Learning Models

TL;DR

This paper investigates how dataset and task overlaps influence the similarity in representations across machine learning models and shows that higher similarity increases vulnerability to certain attacks.

Contribution

It is the first to systematically analyze the causes of representational similarity and its impact on model robustness across different modalities and sizes.

Findings

Dataset and task overlap increase representational similarity.

Greater similarity correlates with higher vulnerability to adversarial attacks.

Combining dataset and task overlap yields the strongest similarity effects.

Abstract

Numerous works have noted similarities in how machine learning models represent the world, even across modalities. Although much effort has been devoted to uncovering properties and metrics on which these models align, surprisingly little work has explored causes of this similarity. To advance this line of inquiry, this work explores how two factors - dataset overlap and task overlap - influence downstream model similarity. We evaluate the effects of both factors through experiments across model sizes and modalities, from small classifiers to large language models. We find that both task and dataset overlap cause higher representational similarity and that combining them provides the strongest effect. Finally, we consider downstream consequences of representational similarity, demonstrating how greater similarity increases vulnerability to transferable adversarial and jailbreak attacks.