Most modern applications in artificial intelligence (AI) rely on centralized data centers to perform the training and inference over large datasets. These data centers are often owned by industrial super-powers, such as Google, Facebook, and Baidu. Individual users or businesses must sacrifice their privacy in order to benefit from the AI services. Distributed (or federated learning) techniques allow learning to be moved from large data centers to distributed edge entities, where users have more control of both their own data and computational resources. In this case, local entities can collaborate to perform learning and inference tasks without explicitly exchanging local raw information with the data center. In this talk, I will first introduce some basic optimization algorithms, such as gradient descent and alternating direction method of multipliers (ADMM), used for learning from decentralized datasets. Then, through the introduction of our recent works on distributed matrix factorization, distributed clustering, and personal authentication etc, we identify several key design challenges and possible ways to address them. The concept of learning from decentralized datasets is consistent with the recent trend in edge computing, where conventional cloud resources are moved towards edge servers or devices. Hence, we will also discuss some novel cross-layered communication or networking algorithms that can be used to reduce the communication overhead under the edge computing framework.
Y.-W. Peter Hongreceived his B.S. degree from National Taiwan University in 1999, and his Ph.D. degree from Cornell University in 2005, both in electrical engineering. He joined the Institute of Communications Engineering and the Department of Electrical Engineering at National Tsing Hua University, Hsinchu, Taiwan, in Fall 2005, where he is now a Full Professor. His research interests include signal processing for sensor networks, UAV communications, distributed learning and optimization, physical layer secrecy, and multiuser wireless communications.