����������

Zhiyu Zhang
Postdoctoral Fellow, Electrical Engineering, Harvard University
Wednesday, Feb. 26 | 10 a.m. | AERO 114

Abstract: Despite the advancements of autonomous systems from decades of engineering, there is always the need to make them even more efficient and reliable. Machine learning holds great potential to achieve this goal, as it can leverage computation and data on an unprecedented scale. An important challenge is thus synergizing these two separate areas, which requires fundamental algorithmic innovations due to the high stakes of interacting with the physical world.

In this talk, I will describe my unique approach to tackle this challenge, specifically from the perspective of adaptive online optimization. This is a major research topic within theoretical machine learning, but my talk will focus on its engineering implications tailored to autonomous systems. The central question is the following: given a “black box” machine learning module, how can we use principled insights from adaptive online learning to improve its efficiency and reliability?

My talk will answer this question in two concrete problems. First, based on [arxiv:2402.02720] (ICML’24) and [arxiv:2410.02561] (in submission), I will demonstrate how to sequentially build trustworthy confidence set predictions on top of an arbitrary point-predicting machine learning model, without explicit statistical assumptions on the nature. Next, based on [arXiv:2405.16642] (NeurIPS’24), I will show that in lifelong reinforcement learning, a theoretically-grounded regularizer can mitigate an intriguing collapse behavior called “loss of plasticity”. These results can be applied to various high-impact modalities of autonomous systems.

Bio: Zhiyu Zhang is currently a postdoctoral fellow in electrical engineering at Harvard University. He obtained his PhD in systems engineering from Boston University, and BEng in mechanical engineering from Tsinghua University. His research centers around the theory and practice of adaptive online learning, which concerns optimal sequential decision making with Bayesian-type prior knowledge. On the application side, he is also excited about various aspects of robotics and automation, especially algorithmic approaches that efficiently utilize large-scale pretraining. He has been recognized by the BU systems engineering outstanding PhD dissertation award, as well as outstanding reviewer awards from NeurIPS, ICML and AISTATS. He also serves as an action editor for the journal TMLR.��