10/08/2026
By Joris Roos

The Department of Mathematics and Statistics invites you to attend a colloquium lecture by our new faculty member Ke Huang.

  • Title: Statistical learning theory for Convolutional Neural Networks (CNNs)
  • Date: Wednesday, Oct. 28
  • Time: 11 a.m. – noon
  • Location: Southwick Hall, Room 350W

Everyone is welcome!

Abstract

Convolutional neural networks (CNNs) have achieved remarkable empirical success, yet a theoretical explanation of why they can be so powerful and how their architecture affects statistical performance remains poorly understood. In this talk, I will present statistical theory for CNNs trained by empirical risk minimization under Lipschitz losses. The framework covers a flexible class of CNN architectures and applies to both regression and classification tasks.

I will discuss non-asymptotic error bounds that characterize how approximation and estimation errors depend on key architectural features of CNNs. I will also show that when the target function lies on a low-dimensional manifold, CNNs can mitigate the curse of dimensionality. Finally, I will illustrate a theory-supported CNN on electrocardiogram (ECG) data, demonstrating strong predictive performance and architectural flexibility compared with existing theoretically supported models.

Visit the Mathematics and Statistics colloquium page.