Seminars & Colloquia Calendar
The topology, geometry, and combinatorics of ReLU neural network functions
Elisenda Grigsby (Boston College)
Date & time: Wednesday, 08 December 2021 at 3:30PM - 4:30PM
Abstract: Feedforward neural networks are a class of parameterized functions that have proven remarkably successful at making predictions about unseen data from finite labeled data sets. They do so even (or rather-especially) in regimes where classical notions of complexity suggest that they ought to be overfitting the training data.
I will describe an ongoing project, joint with Kathryn Lindsey and Marissa Masden, aimed at studying natural topological notions of complexity that we hope will shed light on why overparameterized networks perform so well. Along the way, I will introduce you to some of the beautiful geometry and combinatorics underlying this subject.