Yueling Sun (University of Victoria)

Prototype-based category learning requires the extraction of a central prototype and relies primarily on configural information, namely the spatial relationships among stimulus elements rather than isolated local features. Although deep convolutional neural networks (CNNs) can achieve human-level accuracy in visual recognition, extensive evidence shows that CNNs are biased toward local texture cues and often underuse global structure. It remains unclear whether CNNs use diagnostic information similar to that used by humans in configural prototype-based categorization.

To address this question, this study integrates two complementary interpretability approaches: the ‘Bubbles’ technique to estimate diagnostic visual information in human observers and Grad-CAM to identify image regions guiding CNN decisions. In this talk, I will present the method and preliminary results from a prototype-based category learning task and examine differences in strategy between humans and the CNN model.