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

ANNs that perform well on a Gestalt benchmark from Brain-Score do not support human-like Gestalt processing

Marin Dujmovic¹, Gaurav Malhotra², Jeffrey Bowers¹; ¹University of Bristol, ²State University of New York at Albany

Presenter: Jeffrey Bowers

The BMD2024 behavioural benchmark is designed to test for the Gestalt rule of good continuation that plays an important role in human vision. It came in second place at the 2024 Brain-Score competition aimed at severely testing representative ANNs. Since then, as part of Brain Score, a total of 171 models have been tested on this and three related benchmarks, with the top model performing almost perfectly on all four datasets. Does this suggest that newer models have learned human-like Gestalt rules? We show that the more successful ANNs tend to be the ones largest in scale (both size of the network and the training set), and that the top-performing model performed dramatically worse when the dataset was systematically altered to rigorously test its alignment to humans. This indicates that despite high benchmark scores, these models still do not grasp fundamental principles of human vision, emphasizing the need for controlled experiments to understand their processing mechanisms.

Topic Area: Computational Models of Vision & Visual Cortex