Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Deep Neural Network Models Capture Ventral Stream Predictivity at Cross-Animal Consistency

Josh Wilson¹, Khaled Jedoui Al-Karkari¹, Daniel LK Yamins¹; ¹Stanford University

Presenter: Josh Wilson

The strongest test of a computational model of the brain is whether it can masquerade as another brain. One way to operationalize this test is to predict neural responses in one animal from either a second animal's neural responses or a model's representations, given the same stimuli. If the model-to-brain mapping recovers the same predictive magnitude and hierarchical predictive structure as the brain-to-brain mapping, the model passes as a model of the brain. Here, we evaluate a self-supervised vision transformer (DINOv2) against this standard using large-scale multi-unit recordings spanning V1, V4, and IT from two macaques that viewed the same large set of visual stimuli. We predict each electrode's responses via ridge regression from either the other animal's responses or model layer features. Both mappings recover the same hierarchical organization — early representations best predict V1, intermediate V4, late IT — and the model's peak predictivity approaches cross-animal levels in all three areas. Moreover, the full pattern of cross-area predictivities in brain-to-brain mappings is quantitatively reproduced by the model, demonstrating that a computational model can masquerade as another brain in predicting the ventral visual hierarchy.

Topic Area: Computational Models of Vision & Visual Cortex