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

fMRI and Electrophysiological Recordings are Equally Sensitive for Differentiating Deep Neural Network Models of the Brain

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

Presenter: Daniel LK Yamins

Evaluating how well computational models predict brain responses has become a standard tool for understanding biological vision. Two recording modalities dominate this effort: macaque electrophysiology, which offers fine temporal and spatial resolution across select areas, and human fMRI, which offers whole-brain coverage at coarser spatiotemporal resolution. Given these fundamental differences, it is plausible that the two modalities would yield different model evaluations. Yet whether they agree has never been established. We compare neural alignment scores of 86 models across 10 architecture families on human 7T fMRI and 3 macaque electrophysiology datasets at matched visual areas. The two modalities produce highly similar model rankings: V1 (r = 0.892), V4 (r = 0.908), HighVentral/IT (r = 0.919), replicated across datasets (meta-analytic r = 0.846). Agreement holds across all model layers (r = 0.82 to 0.86), indicating structural convergence. We suggest this reflects a "super-resolution" property: thousands of coarse voxel measurements collectively provide as much information as electrophysiological recordings.

Topic Area: Computational Models of Vision & Visual Cortex