【Team】 Uncertainty Quantification Team
【Date】2026/February/2(Monday) 15:00-16:00(JST)
【Speaker】Talk by Dr. Chao-Kai Chiang ,Project Assistant Professor at the Sugiyama–Yokoya–Ishida Laboratory, The University of Tokyo
Title: Learning with Imperfect Signals: Perspectives from Weakly Supervised Learning and Online Learning
Abstract:
Developing reliable and trustworthy machine learning systems requires both technological advances and solid theoretical understanding for learning with imperfect signals. In this talk, I will present my research path toward this goal from two perspectives: weakly supervised learning and online learning.
I will begin by introducing a unified framework that subsumes the formulations and analyses of seventeen weakly supervised learning scenarios. To demonstrate the applicability of weakly supervised learning, I will then present a case study in reinforcement learning, as well as an ongoing project on learning with label noise.
Imperfect information can also arise from not knowing the future, a central theme in online learning. I will discuss online convex optimization and how to obtain improved performance bounds in gradually changing, yet benign, environments. I will then present recent progress on extending the optimality of Thompson sampling to multi-armed bandits with heavy-tailed reward distributions and an application to large language model routing. Finally, I will conclude with future directions and open challenges in learning with imperfect information.
About the Speaker:
Dr. Chao-Kai Chiang is a Project Assistant Professor at the Sugiyama–Yokoya–Ishida Laboratory, The University of Tokyo. His research focuses on weakly supervised learning, learning with label noise, and multi-armed bandits, with the goal of developing reliable and theoretically grounded machine learning methods for real-world data imperfections.
