Event Date
Speaker: Dogyoon Song, Assistant Professor, Statistics, UC Davis
Title: "Taming Overparameterized Models: OLS Interpolators, Regularization, and Beyond"
Abstract: Recent advances in deep learning research have uncovered the phenomenon of benign overfitting, where models that perfectly interpolate training data can still generalize well, and an implicit bias in gradient descent dynamics, favoring structured solutions linked to implicit regularization. This talk covers two ongoing lines of research addressing the behavior of overparameterized models for high-dimensional statistics. First, we examine the minimum l2-norm OLS interpolator, highlighting key results such as the Frisch-Waugh-Lovell theorem and the leave-one-out formula in overparameterized settings. These results have implications for controlling omitted variable bias and enabling inference. Second, we explore regularized M-estimators and the structures induced by regularization. We revisit classical LASSO analysis, discuss challenges beyond quadratic loss and polyhedral regularizers, and outline ongoing research efforts to tackle these challenges.
Bio: Dogyoon Song is an Assistant Professor of Statistics at UC Davis. Before joining UC Davis, he was a postdoctoral research fellow at University of Michigan. He completed his Ph.D. in EECS at MIT, and earned undergraduate degrees from Seoul National University. His research interests lie at the interface of optimization, statistics, and machine learning with a focus on developing theories and efficient algorithms for data-assisted decision making.
Seminar Date/Time: Thursday October 10, 2024, at 4:10pm
Location: MSB 1147 (Refreshments at 3:30pm, MSB courtyard)