Committee: 

Dr. Dyer, Advisor

Dr. Davenport, Chair

Dr. Muthukumar

Abstract: The objective of the proposed research is to understand and improve machine learning model generalization with data augmentation and optimal transport when the data is subject to distribution shifts. By appealing to a recent connection of optimal transport and the generalization bound, our research aims at unifying the data augmentation and optimal transport that yields the optimal augmentation strategy and robust model calibration. Our theoretical analysis will shed light on when and what data augmentation improves generalization, while the algorithmic design will improve optimal transport's robustness and sampling complexity.