报告题目：To Talk or to Work: Efficient Federated Learning over Mobile Devices
Dr. Miao Pan is an Associate Professor in the Department of Electrical and Computer Engineering at University of Houston. He was a recipient of NSF CAREER Award in 2014. Dr. Pan received his Ph.D. degree in Electrical and Computer Engineering from University of Florida in August 2012. Dr. Pan's research interests include wireless/AI for AI/wireless, deep learning privacy, wireless networks, underwater communications and networking, and cybersecurity. He has published more than 110 papers in prestigious journals including IEEE/ACM Transactions on Networking, IEEE Journal on Selected Areas in Communications, IEEE Transactions on Mobile Computing, and IEEE Transactions on Smart Grid, and around 130 papers in top conferences such as IEEE INFOCOM, ACM MobiSys, ICCV, IEEE ICDM, AAAI, ACM CIKM, IEEE BigData, ICDCS, and IEEE IPDPS. His work won IEEE TCGCC (Technical Committee on Green Communications and Computing) Best Conference Paper Awards 2019, and Best Paper Awards in ICC 2019, VTC 2018, Globecom 2017 and Globecom 2015, respectively. Dr. Pan serves as a Technical Reviewer for many international journals and conferences. He has also been serving as a Technical Organizing Committee for several conferences such as TPC Co-Chair for Mobiquitous 2019, ACM WUWNet 2019, and Technical Program Committee member of several top international conferences, e.g., IJCAI, IEEE INFOCOM, MobiHoc, ICDCS, etc. Dr. Pan is an Editor for IEEE Open Journal of Vehicular Technology and an Associate Editor for IEEE Internet of Things (IoT) Journal (Area 5: Artificial Intelligence for IoT), and used to be an Associate Editor for IEEE Internet of Things (IoT) Journal (Area 4: Services, Applications, and Other Topics for IoT) from 2015 to 2018. Dr. Pan is a member of AAAI, a member of ACM, and a senior member of IEEE and IEEE Communications Society.
In the era of AI and mobile Internet, federated learning (FL) over mobile devices, one of the most anticipated technologies, has fostered numerous intriguing applications and services, while many of them are energy-hungry and delay-sensitive. In this talk, I would like to discuss our proposed energy/delay efficient FL designs over mobile devices. Different from most existing communication efficient FL designs, which regard wireless communications as the bottleneck, we have an interesting observation that under many typical FL settings, the local computing energy/delay is comparable to the communication energy/delay during the FL training process, given the development of high-speed wireless transmission techniques (e.g., 4G LTE, 5G, Wi-Fi 5 or Wi-Fi 6). Thus, the total energy/delay should be computing energy/delay + communication energy/delay over training rounds in FL. To effectively reduce the overall energy/delay of FL over mobile devices, simply reducing local computing/communication energy/delay independently is not enough, and the energy/delay trade-off between "working" (i.e., local computing) and "talking" (i.e., wireless communications) has to be considered. To study this trade-off, we have initially modeled and empirically studied the impacts of advanced local computing control (e.g., DVFS, local iteration number), compression (e.g., gradient quantization, gradient sparsification) and wireless transmission (e.g., transmission rates) strategies on computing and communication energy/delay, formulated the problem of efficient FL over mobile devices, and provided feasible solutions. We have also set up a testbed of FL over mobile devices with Lambda server, Nvidia TX2s/Xaviers and NI USRPs to validate our analysis, and shown promising experimental results.