Similarity in filter properties related to similarity in filter output-connection patterns: analysis using convolutional neural networks

Abstract Neighboring neurons in the cerebral cortex share response properties and project to similar targets. This study clarifies the computational advantage of such organization by examining the relationship between the similarity of output-connection patterns and the similarity of filter properties using a convolutional neural network. After training, the filters in the first layer of the convolutional neural network acquire a variety of structures and develop output-connection patterns to filters in the second layer. Two filters in the first layer with similar output-connection patterns to the second layer have similar degrees of orientation and color selectivity and similar preferred orientations and colors. Thus, these two filters tend to have similar filter properties. This spontaneous emergence of relationships between output-connection patterns and filter properties in a convolutional neural network demonstrates the computational advantage of such an organizational structure in both artificial and biological neural networks.