ECE Seminar* Speaker: Dr. Tharm Ratnarajah, Professor, Institute for Digital Communications, The University of Edinburgh, Edinburgh, U.K.
When: Monday,
October 24, 2022
11:00 AM
-
12:00 PM
Where: Science & Engineering Building, Lester W. Cory Conference Room: Room 213A
Cost: Free
Description: Topic: Federated Learning in Massive MIMO Networks: Convergence Analysis and Communication-Efficient Design
Speaker: Dr. Tharm Ratnarajah, Professor, Institute for Digital Communications, The University of Edinburgh, Edinburgh, U.K.
Location: Lester W. Cory Conference Room, Science & Engineering Building (SENG), Room 213A
Abstract:
In federated learning (FL), model weights must be updated at local users and the base station (BS). These weights are subjected to uplink (UL) and downlink (DL) transmission errors due to the limited reliability of wireless channels. In this paper, we investigate the impact of imperfections in both UL and DL links. First, for a multi-user massive multi-input-multi-output (mMIMO) 6G network, employing zero-forcing (ZF) and minimum mean-squared-error (MMSE) schemes, we analyze the estimation errors of weights for each round. A tighter convergence bound on the modelling error for the communication efficient FL algorithm is derived from the order of O (1/T s2), where s2 denotes the variance of overall communication error, including the quantisation noise. The analysis shows that the reliability of DL links is more critical than that of UL links. The transmit power can be varied in the training process to reduce energy consumption. We also vary the number of local training steps, average codeword length after quantisation and scheduling policy to improve communication efficiency. Simulations with image classification problems on MNIST, EMNIST and FMNIST datasets verify the derived bound and are useful to infer the minimum SNR required for successful convergence of the FL algorithm.
Biography:
Tharm Ratnarajah is currently a Professor in digital communications and signal processing with the Institute for Digital Communications, The University of Edinburgh, Edinburgh, U.K. He has authored or co-authored over 400 publications and holds four U.S. patents. He has supervised 16 PhD students and 21 postdoctoral Research Fellows and raised more than USD 11 million in research funding. His research interests include signal processing and information-theoretic aspects of beyond 5G wireless networks, full-duplex radio, mmWave communications, random matrices theory, and array/radar signal processing.
The Seminars is open to the public free of charge.
*For further information, please contact Dr. Dayalan Kasilingam at 508.999.8534, or by via email at dkasilingam@umassd.edu.
Speaker: Dr. Tharm Ratnarajah, Professor, Institute for Digital Communications, The University of Edinburgh, Edinburgh, U.K.
Location: Lester W. Cory Conference Room, Science & Engineering Building (SENG), Room 213A
Abstract:
In federated learning (FL), model weights must be updated at local users and the base station (BS). These weights are subjected to uplink (UL) and downlink (DL) transmission errors due to the limited reliability of wireless channels. In this paper, we investigate the impact of imperfections in both UL and DL links. First, for a multi-user massive multi-input-multi-output (mMIMO) 6G network, employing zero-forcing (ZF) and minimum mean-squared-error (MMSE) schemes, we analyze the estimation errors of weights for each round. A tighter convergence bound on the modelling error for the communication efficient FL algorithm is derived from the order of O (1/T s2), where s2 denotes the variance of overall communication error, including the quantisation noise. The analysis shows that the reliability of DL links is more critical than that of UL links. The transmit power can be varied in the training process to reduce energy consumption. We also vary the number of local training steps, average codeword length after quantisation and scheduling policy to improve communication efficiency. Simulations with image classification problems on MNIST, EMNIST and FMNIST datasets verify the derived bound and are useful to infer the minimum SNR required for successful convergence of the FL algorithm.
Biography:
Tharm Ratnarajah is currently a Professor in digital communications and signal processing with the Institute for Digital Communications, The University of Edinburgh, Edinburgh, U.K. He has authored or co-authored over 400 publications and holds four U.S. patents. He has supervised 16 PhD students and 21 postdoctoral Research Fellows and raised more than USD 11 million in research funding. His research interests include signal processing and information-theoretic aspects of beyond 5G wireless networks, full-duplex radio, mmWave communications, random matrices theory, and array/radar signal processing.
The Seminars is open to the public free of charge.
*For further information, please contact Dr. Dayalan Kasilingam at 508.999.8534, or by via email at dkasilingam@umassd.edu.
Contact:
ECE: Electrical & Computer Engineering Department 508.999.9164 http://www.umassd.edu/engineering/ece/
Topical Areas: Alumni, Faculty, General Public, SMAST, Staff and Administrators, Students, Graduate, Students, Undergraduate, University Community, College of Engineering, Electrical and Computer Engineering