Title: Convolutional Neural Networks as a Model of Human Intuitive Physical Judgments
Ren Calabro, doctoral student in the Leong Lab, Department of Psychology, University of Chicago
Abstract: Humans reliably infer the physical stability of objects in everyday scenes, yet the cognitive mechanisms underlying these judgments remain unclear. Here, we test whether convolutional neural networks (CNNs), which capture statistical regularities in visual experience without explicit knowledge of physics, can provide a framework for understanding intuitive physical reasoning. We evaluated whether a CNN- specifically, a custom-trained Inception-v4 model trained on the same stimuli presented to human observers, using labels from physics simulations- could predict human stability judgments (N = 500) and the visual features humans attend to when making these judgments. The CNN’s predictions aligned closely with human choices, outperforming ground-truth predictions from physics simulations. Additionally, human eye-gaze patterns correlated with CNN-derived importance maps, suggesting that visual attention is directed toward features statistically predictive of physical outcomes. These findings support the idea that intuitive physics is shaped in part by experience-based visual heuristics and that CNNs can serve as a computational framework for uncovering the attentional and feature-based strategies underlying human judgments.
Time: 03/26/25 3:30 PM
Location: Biopsychological Sciences Building atrium
If you have any questions, requests, and concerns, please contact Nakwon Rim (nwrim [at] uchicago [dot] edu) or Cambria Revsine (crevsine [at] uchicago [dot] edu).