Technical Program

Paper Detail

Paper: PS-2A.28
Session: Poster Session 2A
Location: Symphony/Overture
Session Time: Friday, September 7, 17:15 - 19:15
Presentation Time:Friday, September 7, 17:15 - 19:15
Presentation: Poster
Publication: 2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania
Paper Title: Emergence of Topographical Correspondences between Deep Neural Network and Human Ventral Visual Cortex
Manuscript:  Click here to view manuscript
Authors: Yalda Mohsenzadeh, Caitlin Mullin, Dimitrios Pantazis, Aude Oliva, Massachusetts Institute of Technology, United States
Abstract: Recent computer vision work dissecting information from within the layers of deep neural networks revealed emergence of human-interpretable concepts within these artificial units. In the current study, using representational similarity analysis, we compare convolutional layers of DNNs trained for object and scene recognition (hybrid AlexNet) with regions along ventral visual pathway to ask whether these layers and regions share topographical correspondence. Results reveal the emergence of a brain inspired topographical organization in this hybrid-net, such that layer-units showing strong central-bias were associated with cortical regions with foveal tendencies, and layer-units showing greater selectivity for image boundaries and backgrounds were associated with cortical regions showing strong peripheral preference. The emergence of a categorical topographical correspondence between deepnets and visual regions of interests further strengthens the role of deepnets as models of the inner workings of perceptual networks in the brain.