Causes and Consequences of Representational Similarity in Machine Learning Models

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 dataset and task overlap are positively associated with higher representational similarity across many of our settings, with clear evidence in vision/language classification and weaker trends in language generation experiments. Finally, we consider downstream consequences of representational similarity, showing that greater similarity is associated with increased vulnerability to transferable adversarial attacks in vision models.

Machine learning models often learn similar internal representations, even when they are trained separately. This paper studies why that happens, focusing on two causes: overlap in the training data and overlap in the task the models are trained to solve. Across vision models, language models, and diffusion models, we find that models tend to become more similar when they share more data or more task structure, with the effect being clearest for vision and classification models. We also show one consequence of this similarity: adversarial examples transfer more easily between more similar vision models.