[행사/세미나] 인공지능대학원 전문가 초청 세미나 개최 안내(Dr. Xun Zheng, Software engineer at Waymo, 5/7(금) 10시)
- 인공지능학과
- 조회수25763
- 2021-04-28
*참석 확인을 위해 참여 시 학과명(학번) +이름(인공지능학과20212021 홍길동)으로 접속하여 주시기 바라며, 세미나 참석 활동은 추후 연구수당 평가 기준으로 활용될 수 있습니다.
- Title : Learning DAGs with Continuous Optimization
- Speaker: Dr. Xun Zheng, Software engineer at Waymo (formerly the Google self-driving car project)
- Time : May 7th 10:00 ~ 12:00
- Location: Webex room
- Number: 184 855 0338
- Password: dMw5nVDa4M9
- Link: https://skku-ict.webex.com/skku-ict/j.php?MTID=mfa605b54350017463545943eb4de4b88
1:1 meeting with Dr. Xun Zheong:
After the talk, we provide one-on-one meeting with Dr. Xun Zheng (10 mins each). Please write your name and email address (only school email) if you want to meet him personally until May 2nd.
https://forms.gle/1tzbHWFovXW8XNwG8
Abstract:
Learning the structure of directed acyclic graphs (DAGs, also known as Bayesian networks) from data is an important and classical problem in machine learning, with prominent applications in causal inference, interpretability, robustness, biology, and finance, just to name a few. This is a challenging problem since the search space of DAGs is combinatorial and scales superexponentially with the number of nodes. Existing approaches often rely on various local heuristics for enforcing the acyclicity constraint. By contrast, structure learning for undirected graphical models (e.g. Gaussian MRF) is recognized as a tractable optimization problem nowadays, and achieved huge success in various practical domains such as bioinformatics.
In this talk, we show a first step towards bridging this gap between directed and undirected graphical models. We begin by introducing a fundamentally different strategy for Bayesian network structure learning: We formulate the problem as a purely continuous optimization program over real matrices that avoids the combinatorial constraint entirely. This is achieved by a novel characterization of acyclicity that is not only smooth but also exact. The resulting problem can be efficiently solved by standard numerical algorithms, without imposing any structural assumptions on the graph such as bounded treewidth or in-degree. We show extensions in various settings as well.
Bio:
Xun Zheng is a software engineer at Waymo (formerly the Google self-driving car project), working on planner for autonomous driving. He obtained his PhD in the Machine Learning Department at the Carnegie Mellon University, advised by Eric Xing and Pradeep Ravikumar. His research focuses on probabilistic graphical models, including approximate inference, connections with neural networks, large scale training systems, structure learning and causal discovery. Prior to CMU, he obtained Bachelor from Beihang University, Beijing, China. He is a recipient of the Siebel Scholarship.