Learning Cortical Magnification with Brain-Optimized Convolutional Neural Networks
Florian Mahner, Katja Seeliger, Martin Hebart, Max Planck Institute for Human Cognitive and Brain Sciences, Germany; Umut Güçlü, Donders Institute for Brain, Cognition and Behaviour, Netherlands
Posters 2 Poster
Pacific Ballroom H-O
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
Computational modeling of visual information processing can lead to important new insights about the function of visual cortex. Here we asked whether we can build a proof-of-concept model that implicitly learns known cortical organization principles. We chose cortical magnification, which refers to the fact that more cortical tissue is dedicated to the processing of the foveal as compared to peripheral visual field. We built a brain-optimized convolutional neural network model trained to predict brain activity across twelve retinotopic regions as measured with functional MRI. We treated cortical magnification as a free parameter, using multivariate Gaussian distributions acting on the network's feature representations. Our results demonstrate that cortical magnification can, indeed, be learned implicitly, demonstrating the general feasibility of our computational modeling approach.