Beyond category-supervision: instance-level contrastive learning models predict human visual system responses to objects

Abstract

Anterior regions of the ventral visual stream have substantial information about object categories, prompting theories that category-level forces are critical for shaping visual representation. The strong correspondence between category-supervised deep neural networks and ventral stream representation supports this view, but does not provide a viable learning model, as these deepnets rely upon millions of labeled examples. Here we present a fully self-supervised model which instead learns to represent individual images, where views of the same image are embedded nearby in a low-dimensional feature space, distinctly from other recently encountered views. We find category information implicitly emerges in the feature space, and critically that these models achieve parity with category-supervised models in predicting the hierarchical structure of brain responses across the human ventral visual stream. These results provide computational support for learning instance-level representation as a viable goal of the ventral stream, offering an alternative to the category-based framework that has been dominant in visual cognitive neuroscience.