计算机学院系列讲座菁英论坛第37期——Optimization of discrete optimization in machine learning

报告题目(Title)：Optimization of discrete optimization in machine learning

时间(Date & Time)：2024.11.29; 3-4pm

地点(Location)：理科一号楼1131（燕园校区） Room 1131, Science Building #1 (Yanyuan) 

主讲人(Speaker)：Dianbo Liu（刘钿渤）

邀请人(Host)：Shanghang Zhang（仉尚航）

报告摘要(Abstract)：

Discrete representations play a crucial role in many deep learning architectures, yet their non-differentiable nature poses significant challenges for gradient-based optimization. To address this issue, various gradient estimators have been developed, including the Straight-Through Gumbel-Softmax (ST-GS) estimator, which combines the Straight-Through Estimator (STE) and the Gumbel-based reparameterization trick. In this talk, we share several strategies recently developed in our team to improve efficiency of discrete optimization and demonstrate the usage of discrete representation in manipulating behaviors of language models.

主讲人简介(Bio)：

Dianbo Liu is leader of Cognitive AI for Science team (CogAI4SCI.com) and assistant professor at National University of Singapore. Before starting CogAI4Sci team, Dianbo Liu was a group leader at the Broad Institute of MIT and Harvard. Prior to the Broad Institute, Dianbo worked as a postdoctoral researcher with Prof. Yoshua Bengio (a Turing Award winner) and led the Humanitarian AI team at the Mila-Quebec AI Institute. This followed his fellowship training and studies in medical informatics at Harvard University. Dianbo earned his PhD from the University of Dundee, Scotland, under the supervision of Prof. Timothea Newman. During his doctoral studies, he received the Vest Scholarship from the Massachusetts Institute of Technology (MIT) and was a special graduate student at the MIT Computer Science and Artificial Intelligence Lab. Dianbo also co-founded two start-ups, "GeneTank" and "SecureAILabs," to advance AI applications in biomedical sciences during his training.

欢迎关注计算机学院微信公众号，了解更多讲座信息！

北京大学计算机学院