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

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Scaling Behavioral Probes of Face Processing: Inversion and Thatcher Effects in Deep Networks

Srijani Saha¹, Talia Konkle¹, George A. Alvarez¹; ¹Harvard University

Presenter: Srijani Saha

People recognize faces across dramatic variations in appearance—changes in viewpoint, lighting, age, makeup, and even caricaturization. Holistic processing—the contextual interaction between features producing a unified representation where the whole exceeds the sum of its parts—is the dominant proposed mechanism underlying this ability. Yet no image-computable account broadly explains the behavioral signatures of holistic processing, including the Inversion Effect (degraded recognition for inverted faces) and the Thatcher Effect (heightened sensitivity to feature distortions in upright versus inverted faces). Here we developed a large-scale Thatcher stimulus set drawn from VGGFace2 (75 identities × 15 images × 4 conditions) and two complementary geometric measures that independently quantify Inversion and Thatcher effects. We evaluated these probes across deep neural networks varying in architecture (CNNs, ViTs) and training objective (face identification/discrimination versus general object classification/discrimination). Consistent with prior work, both Inversion and Thatcher effects emerged in the output embeddings of face-trained models regardless of architecture, but were absent in general-purpose vision models. These results establish a suite of image-computable behavioral probes deployable on both humans and models, enabling precision psychophysics studies and fine-grained model–human alignment analyses.

Topic Area: Computational Models of Vision & Visual Cortex