How is scene recognition in a convolutional network related to that in the human visual system?

Abstract

This study is an analysis of scene recognition in a pre-trained convolutional network, to evaluate the information the network uses to distinguish scene categories. We are particularly interested in how the network is related to various areas in the human brain that are involved in different modes of scene recognition. Results of several experiments suggest that the convolutional network relies heavily on objects and fine features, similar to the lateral occipital complex (LOC) in the brain, but less on large-scale scene layout. This suggests that future sceneprocessing convolutional networks might be made more brain-like by adding parallel components that are more sensitive to arrangement of simple forms.