Humans and convolutional neural networks prioritize similar visual features in intuitive physics judgments

Humans reliably infer complex physical relationships between objects in everyday scenes, yet the mechanisms underlying these judgments remain unclear. We explored whether convolutional neural networks (CNNs) can approximate intuitive physical reasoning by capturing statistical regularities in visual experience. We trained a CNN (Inception-v4) to predict tower stability and tested how well its outputs aligned with human judgments (N = 500). CNN predictions more closely matched human judgments (r = 0.718, p < 0.001, accuracy = 81%) than ground-truth predictions from physics simulations (r = 0.406, p = 0.002, accuracy = 68%), suggesting that both CNNs and humans rely on visual heuristics. Eye-tracking data revealed that CNN importance maps overlapped significantly with human gaze patterns, indicating shared attention to features statistically predictive of physical outcomes in intuitive physical judgments. Our findings show that CNNs trained on visual data capture perceptual cues used in human intuitive physics, highlighting their value as models of heuristic reasoning.