Proceedings of the Annual Meeting of the Cognitive Science Society

Humans show systematic biases in geometric perception: shapes with higher regularity (symmetry, right angles, parallel sides) and simplicity (fewer elements) are easier to recognize. These biases have been difficult to replicate in neural networks, leading to the hypothesis that the human visual system uniquely employs a "geometric language of thought''---discrete symbolic machinery for representing shapes---that cannot be reproduced by neural networks. We show that neural networks trained via meta-learning can reproduce these biases without symbolic primitives or massive pretraining. Training on concepts with fixed complexity yields human-like regularity sensitivity, whereas training on concepts with varying complexity yields simplicity sensitivity. This dissociation suggests that meta-learning induces priors that exploit the most diagnostic dimension of variation in the training environment. More broadly, our results suggest that these perceptual biases can arise from the statistics of the training environment itself, rather than being innately endowed.