Can neural networks model the human perception of geometric shapes?

Artificial neural networks achieve impressive success in many vision tasks. Nevertheless, previous work has suggested that their representation of geometric shapes only partially captures the representations found in humans, who additionally require a symbolic "language of geometry". In this study, we evaluate the progress in this area by systematically comparing human performance in three geometric shape perception tests, with the predictions of recent Convolutional Neural Networks (CNNs) and Vision Transformers, varying in size, training datasets and training methods, to recognize and process geometric shapes and compare these representations to the ones found in humans. In two tasks probing quadrilateral and geometric shape perception, introduced by Sable-Meyer et al. (2022, 2026), recent neural networks exhibit representations similar to humans and predict human performance better than a symbolic model. In a third task, involving geometric drawings, the networks still fell short of human performance. Across the three experiments, the main factor driving similarity between human and neural network representations was the size of the training dataset, rather than a specific architecture or the number of parameters of the model. Together, our results show that a massive scaling-up of the training data helps neural networks approximate human-like representations of geometric shapes, but that they still fall short of fully capturing the human representations of geometric shape.