Spontaneous segregation of visual information between parallel streams of a multi-stream convolutional neural network

TL;DR

This study investigates how visual information naturally segregates into different streams within a multi-stream convolutional neural network, mirroring biological visual processing, and finds that color and shape information tend to segregate.

Contribution

The paper demonstrates spontaneous segregation of visual features like color and shape in a multi-stream CNN, reflecting biological parallel processing in the primate brain.

Findings

Color information segregates from shape in most model instances.

Deletion of color stream reduces recognition of inanimate objects.

Deletion of shape stream reduces recognition of animate objects.

Abstract

Visual information is processed in hierarchically organized parallel pathways in the primate brain. In lower cortical areas, color information and shape information are processed in a parallel manner, while in higher cortical areas, various types of visual information, such as color, face, animate/inanimate, are processed in a parallel manner. In the present study, the possibility of spontaneous segregation of visual information in parallel streams was examined by constructing a convolutional neural network with parallel architecture in all of the convolutional layers. The results revealed that color information was segregated from shape information in most model instances. Deletion of the color-related stream decreased recognition accuracy in the inanimate category, whereas deletion of the shape-related stream decreased recognition accuracy in the animate category. The results suggest that properties of filters and functions of a stream are spontaneously segregated in parallel streams of neural networks.