报告题目:  From Reading to Do to Reading to Learn

主讲人:  Prof. Victor Zhong  University of Waterloo, Microsoft Research

报告时间： 1月15日 10:00 —— 11:30

报告地点: 王选计算机研究所 106报告厅

报告摘要： In this talk, I describe recent work on learners that adeptly generalize to new problems and environments by reading problem descriptions. Prior work in machine learning show that models struggle to learn efficiently and generally from sparse, delayed reward signals. In light of this, we investigate the suitability of natural language, given its richness and compositional nature, as a means of efficient and generalized learning. First, we discuss a class of reading methods that read natural language instructions and manuals to generalize to new instructions and new environment dynamics. Second, we explore emerging ideas in Reading to Learn using problem descriptions in 1) sim-to-real methods that transfer to real-world applications by combining pretrained large language models with verbalization of environments 2) shaping policies using naturally existing, abundantly available manuals 3) discovery of manuals for automated curriculum learning. Through this talk, we advocate for using language as a vehicle for efficient and generalized learners, by enabling them to autonomously gather knowledge for formulating comprehensive plans in real-world scenarios.

个人简介: Victor Zhong is an Assistant Professor at the University of Waterloo and a Faculty Member at the Vector Institute. His research is at the intersection of natural language processing and machine learning, and aims to teach machines to read natural language specifications to generalize to new problems. Victor has a Ph.D. from the University of Washington, a M.Sc. from Stanford University, and a B.A.S. from the University of Toronto. He has worked as a researcher at Meta AI Research, Microsoft Research, Google Brain, and was a founding member of Salesforce Research. His research has been awarded the Apple AI/ML Fellowship and Outstanding Paper Award at EMNLP. His work has also been featured by Quanta Magazine, Wired, VentureBeat, MIT Technology Review, Fast Company, and TechCrunch.