报告人: Dr. Kun Yuan, DAMO Academy, Alibaba (US) Group

时  间：1月13日（周四）上午10:00-11:30

报告线上链接：https://meeting.tencent.com/dm/FU5r65QrigPs

腾讯会议号：210-386-097

主持人：文再文 教授

报告内容提纲：

Decentralized algorithms, in which each computing node communicates only with its neighbors, are

more communication-efficient and robust to node failures. Decentralized algorithms are widely used in wireless signal processing, control, and robotics, but they have profound applications in accelerating large-scale deep training problems recently. In this talk, we will first review existing distributed training methods and their communication overhead, and explain how decentralized optimization can help accelerate deep training. Next, we will explore what topology shall we organize all nodes to balance communication efficiency and convergence performance, how to develop algorithms to fit into large-batch training, and how to handle data heterogeneity across all local nodes. These results can make decentralized algorithms much more efficient and practical for deep training. Finally, we will introduce BlueFog, an open-source GitHub repo that we build, to help researchers quickly deploy their newly-developed decentralized methods in deep learning.

报告人简介:

Dr. Kun Yuan received his Ph.D. degree in the Electrical and Computer Engineering at University of California, Los Angeles (UCLA) in 2009. After that, he joined DAMO Academy, Alibaba (US) Group as a research scientist. He was a visiting researcher in EPFL and Microsoft Research Redmond in 2018. Dr. Yuan was the recipient of the 2017 IEEE Signal Processing Society Young Author Best Paper Award, and the 2017 ICCM Distinguished Paper Award. His research mainly focuses on the theory, algorithms, and applications in optimization, signal processing, and machine learning. 

欢迎广大师生光临！