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

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Object symmetry is a prevalent environmental feature, yet its abstract nature poses challenges for traditional computer vision algorithms.
• Understanding how biological and artificial systems perceive symmetry is crucial for advancing both fields.

Purpose of the Study:

• To investigate whether Deep Neural Networks (DNNs) trained on natural images develop symmetry detection capabilities similar to the human visual system.
• To identify the specific layers within DNNs responsible for processing visual symmetry.

Main Methods:

• A DNN was trained on a dataset of natural environmental images.
• The trained DNN was subsequently tested using object-free random-dot images possessing varying degrees of symmetry (1, 2, and 4 axes).
• Symmetry discriminability was analyzed across different layers of the DNN.

Main Results:

• Symmetry coding was minimal in the initial layers of the DNN.
• The most significant symmetry discrimination was observed in the FC6 layer, a fully connected layer.
• This layer's performance characteristics align with the human lateral occipital complex (LOC), a brain region involved in visual processing.

Conclusions:

• Feedforward DNNs trained on natural images exhibit a form of visual processing homologous to the human brain's extended visual hierarchy.
• The findings suggest that DNNs can serve as valuable models for understanding human visual perception, particularly in the domain of symmetry detection.