Time: 4:30-5:30 p.m.
Date: Monday, February 2, 2026
Speaker: Jiawei Ge
Title: Toward Reliable AI: The Statistical Foundations of Out-of-Distribution Generalization
Abstract: Modern machine learning systems are often deployed in environments that differ from their training data, making out-of-distribution (OOD) generalization a central challenge for reliable AI. In this talk, I study the statistical foundations of learning under distribution shift, focusing on the widely encountered setting of covariate shift, where input distributions change but the labeling mechanism remains stable. I show that in the well-specified regime, classical maximum likelihood estimation (MLE) is minimax-optimal for OOD generalization, whereas under model misspecification, appropriately weighted likelihood methods can outperform MLE and achieve minimax optimality in certain scenarios. I then turn to a modern learning regime, in which neural networks can perfectly interpolate the training data in many ways, yet exhibit vastly different generalization behavior. I propose a simplicity principle that explains why some solutions generalize reliably and establish theoretical guarantees for regularized MLE in two distinct regimes.
Bio: Jiawei Ge is a fifth-year Ph.D. candidate in Operations Research and Financial Engineering at Princeton University, advised by Professors Jianqing Fan and Chi Jin. Her research bridges statistical foundations and real-world applications, focusing on developing theoretical guarantees and efficient algorithms that turn vast, heterogeneous data into trustworthy AI systems. Her work spans unsupervised pretraining, out-of-distribution generalization, benchmarking large language models, network modeling, and uncertainty quantification. Her contributions have been recognized in premier machine learning conferences as well as leading statistical journals.
