Professor Wenchao Li Earns NSF CAREER Award
by A.J. Kleber
The National Science Foundation has presented Assistant Professor Wenchao Li (ECE, SE) with a prestigious CAREER Award for a new project which has the potential to transform the development of AI systems.
Titled “Specification-Guided Imitation Learning,” Professor Li’s project aims to overcome fundamental challenges in Imitation Learning (IL), a powerful machine learning approach which allows AI agents to learn from “expert demonstrations,” such as videos or other recorded data of a human performing a particular task or activity. However, Professor Li notes that “in practice, human demonstrations can be inadequate, partial, imperfect, environment-specific, or suboptimal.” To address this, he proposes a novel approach which adds a “guiding framework” to IL.
One of the major areas of concern in ML is providing effective safety parameters for technologies like autonomous vehicles. Relying exclusively on demonstrated driving practices could leave critical weaknesses in such a system, because training on exclusively safe and successful scenarios fails to provide the context to recognize unsafe situations; AI systems are not well equipped to generalize to broader conditions than their training covers. At the same time, providing demonstrations addressing every potential safety risk would be highly impractical.
To address this shortcoming, Li proposes expanding the “expert inputs” provided in IL to include formal specifications for a given task, to further define the expectations and requirements for the “learner.” By utilizing a complimentary, mathematically precise framework to clarify and provide context, Li plans to vastly improve the overall effectiveness of IL training in general. “With our new approach, we envisage that it will lead to more efficient learning algorithms, AI systems that are better aligned with users’ intents, and ultimately safer and more trustworthy autonomous systems,” Li affirms.
Professor Li leads the aptly-named Dependable Computer Laboratory; his research interests include trustworthy AI, cyber-physical systems and design automation. He was a Hariri Institute for Computing Junior Faculty Fellow and Peter J. Levine Career Development Professor from 2018-2021, and received a Hariri Research Incubation Award in 2018 as well. Before joining BU ECE in 2016, he worked as a Computer Scientist at SRI International, Menlo Park. He earned his Ph.D. in Electrical Engineering and Computer Sciences in 2013 at the University of California, Berkeley, where his research received the ACM Outstanding Ph.D. Dissertation Award in Electronic Design Automation, and the Leon O. Chua Award for outstanding achievement in the area of nonlinear science.