Speaker

Changlong Wu

Abstract

Online learning is a fundamental machine learning paradigm in which the learner receives data sequentially, continually updating a model to predict future observations. Real-world scenarios, from financial markets to recommendation systems, however, often exhibit complex data characteristics, mixing structured patterns with unpredictable shifts. The complexity of such systems cannot be adequately captured by traditional methods, which typically assume that data is either purely stochastic or purely adversarial.

In this talk, I introduce a novel theoretical framework that bridges this gap by accommodating general "hybrid" data generation processes. To illustrate the core ideas underlying the framework, I will focus on a specific setting where features are generated by an unknown nonstationary process while labels are generated adversarially. The complexity of feature generation is quantified by the number of different distributions it can change to over time. I will present a general prediction strategy for this scenario, achieving tight regret guarantees with respect to both the complexities of the model class and the data generation. The core idea is based on a novel concept of a stochastic sequential cover, which also applies in broader contexts. I will then present computationally efficient prediction strategies using a novel follow-the-perturbed-leader-type algorithm with self-generated dummy samples, demonstrating the practical feasibility of our framework. Finally, I will briefly discuss several future directions, such as learning fault-tolerant models and characterizing hallucinations in large language models. Collectively, these efforts advance the foundational principles of trustworthy machine learning.

Bio

Dr. Changlong Wu is a Visiting Assistant Professor in the Department of Computer Science at Purdue University. Previously, he was a Postdoctoral Research Associate at the Center for Science of Information (CSoI) at Purdue University from 2021 to 2023. He earned his Ph.D. in Electrical Engineering from the University of Hawai‘i at Mānoa in 2021. His research focuses on theoretical machine learning, statistics, and the foundations of AI, with a particular emphasis on online learning, sequential decision-making, and robust learning under uncertainty. His work has been published in top-tier venues, including COLT, NeurIPS, ICML, ICLR, and IEEE-TIT, with recognition for an Oral presentation at NeurIPS 2022 and a Spotlight at ICML 2024.