Keynote Speaker: Weiran Wang, Ph.D. - 2026 Prospective Student Visit Day and Graduate Research Symposium

Title: Tuning Distributionally Robust Learning via Bilevel Minimax Optimization

Abstract: Deep learning models increasingly rely on Distributionally Robust Optimization (DRO) to combat subpopulation shifts and spurious correlations. However, advanced DRO methods depend on sensitive hyperparameters that traditionally demand computationally costly grid searches. We propose a framework that treats the robustness mechanism itself as a learnable component, continuously tuning these parameters on held-out validation data directly within the training loop.

Because robust training is inherently a minimax game, optimizing its hyperparameters creates a complex nested bilevel minimax structure. To solve this, we introduce a novel penalty-based first-order method that operates efficiently without relying on restrictive lower-level strong convexity assumptions. Furthermore, we provide a generalization guarantee for our method, which dynamically discovers the bias-variance trade-off on the fly, bypassing the discretization errors and massive computational overhead of grid searches.

This talk covers joint work with Yiyang Shen (PhD@Informatics), Qihang Lin (Tippie College of Business), and AI.

Bio: Weiran Wang is an Assistant Professor in the Department of Computer Science (with joint affiliations in AMCS and Informatics) at the University of Iowa. Prior to joining academia, he spent seven years in industrial research positions. He received his PhD from the University of California, Merced and was a postdoctoral researcher at the Toyota Technological Institute at Chicago (TTIC). His research takes an algorithmic approach to machine learning, focusing on multimodal representation learning, generative models, optimization for machine learning, and their applications to speech processing and large-scale language models.

2026 Prospective Student Visit Day and Graduate Research Symposium details here