
Towards a Geometric Theory of Deep Learning
Govind Menon of Brown University, speaking at the Institute for Advanced Study, lays out a mathematical framework for understanding deep learning through the deep linear network (DLN), a phenomenological model introduced by Arora, Cohen and Hazan in 2017. Menon presents results on training dynamics under stochastic gradient descent, showing unexpected connections to minimal surfaces, geometric invariant theory, and random matrix theory. He frames the mathematical core of deep learning as a function approximation problem and sketches a conceptual picture of what he calls 'true' deep learning, drawing on joint work with Nadav Cohen, Kathryn Lindsey, Alan Chen, Tejas Kotwal, Zsolt Veraszto and Tianmin Yu. Delivered to a research mathematics audience at Simonyi Hall, the talk moves through dense technical material connecting neural network theory to established areas of pure mathematics, aimed at listeners already familiar with analysis and geometry.