The canonical deep neural network as a model for human symmetry processing

A key property of our environment is the mirror symmetry of many objects, although symmetry is an abstract global property with no definable shape template, making symmetry identification a challenge for standard template-matching algorithms. We therefore ask whether Deep Neural Networks (DNNs) trained on typical natural environmental images develop a selectivity for symmetry similar to that of the human brain. We tested a DNN trained on such typical natural images with object-free random-dot images of 1, 2, and 4 symmetry axes. Symmetry coding was negligible in the earliest DNN layers. The strongest discriminability occurred in the first fully connected layer, FC6, plausibly analogous to the human lateral occipital complex (LOC), matching many structural properties of human symmetry processing. These results support the homology between the feedforward DNN trained on natural images and the global processing of the extended visual hierarchy as it has evolved in the human brain.