Physical Society Colloquium

Opening the neural network black box

Brice Ménard

Department of Physics and Astronomy
John Hopkins University

Neural networks can perform complex tasks but how and what they learn during their training phase has remained mysterious for a long time. I will present recent results providing valuable insights into this question and showing how to make sense of the network weights. I will explain how to extract weight summary statistics and how we can use them to inspect what has been learned as well as create new networks that do not need to be trained. Finally, I will explain how these new concepts allow us to better understand biological brains and I will show recent results exploring the visual cortex of mice and humans using calcium imaging and fMRI data.