Title: Generalization in Dynamic Environments: Foundations and Algorithms

Date: Wednesday, April 1st, 2026

Time: 10:00am - 12:00pm EDT

Location: Coda C1215 Midtown

Teams link: Tyler LaBonte Thesis Defense | Meeting-Join | Microsoft Teams

Tyler LaBonte

Machine Learning Ph.D. Candidate

School of Industrial and Systems Engineering

Georgia Institute of Technology

Committee

1. Dr. Vidya Muthukumar (Advisor), School of Electrical and Computer Engineering and School of Industrial and Systems Engineering, Georgia Institute of Technology 

2. Dr. Jacob Abernethy (Co-advisor), School of Computer Science, Georgia Institute of Technology

3. Dr. Steve Mussmann, School of Computer Science, Georgia Institute of Technology

4. Dr. Justin Romberg, School of Electrical and Computer Engineering, Georgia Institute of Technology

5. Dr. Surbhi Goel, School of Engineering and Applied Science, University of Pennsylvania

Abstract

Despite the success of modern machine learning models in a vast number of data-rich domains, an incomplete understanding of their robustness limits deployment in high-consequence scenarios. In particular, generalization often suffers as real-world conditions deviate from training environments and the strict assumptions of classical theory no longer apply. These statistical phenomena precipitate serious impacts in machine learning applications: for example, medical diagnosis algorithms can fail when deployed in new hospitals, and recidivism prediction is known to exacerbate bias against demographic minorities.

This thesis represents a joint theoretical and empirical investigation of generalization in dynamic environments. Its core tenet is that an understanding of the underlying mechanisms of generalization phenomena can elicit elegant methodological improvements. Therefore, we place equal emphasis on foundations -- mathematical or experimental analyses which divulge fundamental insights into model behavior -- and algorithms -- adaptive methods which ensure model trustworthiness via rigorous benchmarking or theoretical guarantees.

We begin the talk by summarizing proposal work in spurious correlations: patterns which are predictive of the target class in the training dataset but irrelevant to the true classification function, and task shift: wherein the objective function at test-time differs from the training objective. The majority of the talk will concern new work on a theoretical characterization of the spurious feature learning dynamics of a two-layer ReLU neural network in a Boolean data setting. We show that a linear spurious correlation is learned exponentially fast and greatly suppresses the learning of a quadratic core feature. To our knowledge, our results comprise the first end-to-end theoretical analysis of spurious feature learning in neural networks outside the lazy (linearized) regime.