Presentations

Neural Networks Can Automatically Adapt to Low-Dimensional Structure in Inverse Problems

September 04, 2025

Talk, Brigham Young University Applied Analysis Seminar, Provo, Utah

Abstract: Machine learning methods are increasingly used to solve inverse problems, wherein a signal must be estimated from few measurements generated via a known acquisition procedure. While approaches based on neural networks perform well empirically, they have limited theoretical guarantees. Specifically, it is unclear whether neural networks can reliably take advantage of low-dimensional structure shared by signals of interest — thus facilitating recovery in settings where the signal dimension far exceeds the number of available measurements. In this talk, I will present a positive resolution to this question for the special case of underdetermined linear inverse problems. I will show that, when trained with standard techniques and without explicit guidance, deep linear neural networks automatically adapt to underlying low-dimensional structure in the data, resulting in improved robustness against noise. These results shed light on how neural networks generalize well in practice by naturally capturing hidden patterns in data.

Finding Low-Rank Functions Using Linear Layers in Neural Networks

February 14, 2023

Talk, University of Chicago Computational and Applied Mathematics Student Seminar, Chicago, Illinois

A fundamental question in the theory of neural networks is the role of depth. Empirically it is widely known that deeper networks tend to perform better than shallow ones. However, the reasoning behind this phenomenon is not well understood. In this talk I will discuss the role of depth in the simplified case where most of the layers have a linear activation. Specifically, the regularization associated with training a neural network with many linear layers followed by a single ReLu layer using weight decay is equivalent to a function-space penalty that encourages the network to select a low-rank function, i.e. one with small active subspace.