(A) simulation study on the emergence of visual symmetry perception using an artificial neural network model

Abstract

The ability to perceive the reflectional symmetry of visual objects is crucial for hunting and mating behavior of various animal species. The symmetry-selective population response in area V4 and the lateral occipital complex is thought to provide a basis for symmetry perception, yet whether symmetry-selectivity can arise in individual neurons remains elusive. Here, using a biologically inspired deep neural network model, we show that selective tuning to symmetry can arise at the level of single neurons, solely from the visual experience of natural images. By measuring the responses of individual neurons in a network to a visual stimulus with varying degrees of symmetry, we found neurons showing monotonically increasing or decreasing responses with respect to the degree of symmetry. Those symmetry-selective neurons could reproduce the characteristics of human symmetry perception, such as the distance effect and the orientation effect, referring to the anisotropy of symmetry-sensitivity for different symmetry axis orientations. To investigate the developmental mechanism of symmetry-selective neurons, we conducted a simulation in which networks were trained on natural image datasets with controlled degrees of symmetry. We observed a significant positive correlation between the degree of symmetry of the training dataset and the number of symmetry neurons arising after training, and the orientation effect could be modulated by a similar procedure, showing how the statistics of natural images can be reflected in the behavioral characteristics of symmetry-selectivity. Overall, our results provide a prediction of a neuronal basis of symmetry perception and suggest a possible mechanism for the spontaneous emergence of symmetry-selectivity based on the experience of natural images, even in the absence of training on symmetry detection.