Shape-selective processing in deep networks: integrating the evidence on perceptual integration

Understanding how deep neural networks resemble or differ from human vision becomes increasingly important with their widespread use in Computer Vision and as models Neuroscience. A key aspect of is shape: we decompose the visual world into distinct objects, cues to infer 3D geometries, can group several object parts a coherent whole. Do shape objects similarly when they classify images? Research on this question has yielded conflicting results, some studies showing evidence for selectivity networks, while others demonstrated clear deficiencies. We argue that these conflicts arise differences experimental methods: whether custom images which only features are available, different compete, image pairs vary along feature dimensions, large sets assess representations overall. Each method offers different, partial view processing. After comparing advantages pitfalls, propose two hypotheses reconcile previous results. Firstly, sensitive local, but not global shape. Secondly, higher layers discard information lower to. test by network natural silhouettes local degraded. The results support both hypotheses, networks. Purely feed-forward convolutional unable integrate globally. In contrast, residual recurrent connections show weak This motivates further research architectures perceptual integration.