A Hippocampal-Entorhinal System Inspired Model for Visual Concept Representation

Visual concepts play a significant role in human cognition. The question of how past experience guides the formation of visual concept system has been contentious for decades. There have been exciting recent progresses on concept learning and visual understanding. Despite the great advances, the previous models ignored the manner of visual concepts organization and representation in the brain. The semantic concepts in these models are not grounded on concrete visual representations. In this paper, we propose a novel framework named visual concept space model (VCSM) by drawing inspiration from the hippocampal-entorhinal system. Specifically, we extend the role of hippocampal-entorhinal system from spatial navigation to visual concept space. The proposed model provides a spatial representation format for visual concepts. The semantic concepts in the VCSM are explicitly built on visual representations. Once trained, VCSM can infer the visual attributes of input images and reconstruct diverse images from given concepts. Comprehensive experimental results on 3D Chairs and Extended Yale Face Database B demonstrate the effectiveness of the proposed model.