10/23/25 Ginni Strehle: Highlighting differences between human face perception and deep neural networks
CCN brown bag
Ginni Strehle
Graduate Student
Date: Thursday, October 23, 2025
Time: 12:10- 1:00 pm
Location: 316 Wilson Hall
Highlighting differences between human face perception and deep neural networks
Human face perception is tuned to subtle differences. What aspects of a face can alter our perception of similarity the most? We explored this question using a 3D model of face appearance, which represented the shapes and textures of real people in two separate principal component spaces. Specifically, we wanted to understand how shape and texture affected the perceived similarity of faces by human observers. Further, could these patterns of human perception be accounted for by deep neural networks (DNNs) trained on face identification? Previous studies have shown that DNN models can predict human perceptual judgments regarding the similarity between a variety of different faces. However, these studies did not examine how shape and texture independently affect face perception. We investigated this question across two experiments. In experiment 1, participants were presented with a target face and two alternate faces that differed in either shape or texture, and indicated which alternate face differed more from the target. Human judgments were compared to the responses evoked by the shape- and texture-altered faces and their corresponding targets in penultimate layers of DNNs. In experiment 2, participants adjusted a face’s shape and/or texture until it appeared to depict an identity that was just noticeably different from that of a reference face. We compared human adjustment thresholds to the representational distances between the altered faces in penultimate layers of DNNs. Across both experiments, we found that human similarity judgments were more influenced by shape information. This contrasted with DNNs, which appeared to be more influenced by texture. Our results indicate that even though DNNs can predict human judgments of face similarity to a modest degree, DNNs do not appear to be as sensitive to shape information as human observers.