Events

MaD Seminar: Opening the neural network black box

Speaker: Brice Menard

Location: 60 Fifth Avenue, Room 150

Date: Thursday, February 23, 2023

I will present a simple point of view allowing us to make sense of the weights in a trained neural network. I will show how to characterize what has been learned, extract quasi-sufficient summary statistics, and use them to generate new networks performing well without any training. I will show that the symmetry group of neural networks are layer-based rotations. When taken into account network weights always converge to the same solution. I will illustrate these results using standard classification tasks on CIFAR-10 and ImageNet and I will introduce a model that captures all these properties. Finally, I will show that most of the stochasticity inherent to neural networks and their training is largely negligible. Collaborators: F. Guth, S. Mallat & G. Rochette.