Allen M. Chen¹ (), Ramanujan Srinath¹, Charles E. Connor¹; ¹Johns Hopkins University

Studies of neural coding in intermediate and higher level stages of the ventral/object pathway of primate visual cortex have revealed that neurons represent 2D and 3D shape in terms of object fragments, their geometric properties, and their object-relative spatial configuration. However, almost nothing is known about the neural algorithms that generate this information from lower level inputs. Here, we used analysis of visual neural network processing to derive hypotheses about these algorithms in area V4, an intermediate stage where individual neurons encode orientation, curvature, and object-relative positions of 2D contour fragments, 3D surface fragments, and 3D medial axis fragments. Our aim is to test these algorithms with experiments in V4, V2, and V1, to discover how this geometric information about object fragments is generated from the 2D Gabor-like filter signals that tile visual space in area V1. We have analyzed AlexNet, a visual network trained on the ImageNet database. Previous work has shown that V4 responses are most closely modeled by layer 3 of such convolutional vision networks. We analyzed how layer 3 response patterns depend on differential inputs from convolutional neurons in layer 2 (conv2) modulated through, max pooling, rectification, and connection weight patterns to the layer 3 neuron. In turn, we analyzed how the responses of these conv2 neurons depended on differential inputs from layer 1 Gabor-like filters (conv1). One hypothesis emerging from these analyses is that V4 neurons that encode 3D shape fragments should be driven almost entirely by Gabor-like signals for achromatic contrast and low spatial frequencies. In contrast, V4 neurons selective for 2D shape should be driven more strongly by signals for chromatic contrast. Our initial tests of this hypothesis contrast the sensitivity of 3D- vs. 2D-responsive neurons in V4 to achromatic vs. chromatic contrast and low vs. high spatial frequencies.