A Learning Approach for the Design of Massive MIMO and Intelligent Reflecting Surface Systems

Wei Yu received the B.A.Sc. degree in Computer Engineering and Mathematics from the University of Waterloo Canada in 1997 and M.S. and Ph.D. degrees in Electrical Engineering from Stanford University, U.S.A., in 1998 and 2002, respectively. Since 2002, he has been with the Electrical and Computer Engineering Department at the University of Toronto, Canada, where he is now Professor and holds a Canada Research Chair (Tier 1) in Information Theory and Wireless Communications.

Prof. Wei Yu is a Fellow of IEEE, a Fellow of the Canadian Academy of Engineering, and a member of the College of New Scholars, Artists and Scientists of the Royal Society of Canada. He received the IEEE Marconi Prize Paper Award in Wireless Communications in 2019, the IEEE Communications Society Award for Advances in Communication in 2019, the IEEE Signal Processing Society Best Paper Award in 2017 and 2008, the Journal of Communications and Networks Best Paper Award in 2017, and the IEEE Communications Society Best Tutorial Paper Award in 2015.

Prof. Wei Yu has been elected and serves currently as the President of the IEEE Information Theory Society in 2021.

In this talk, we explore the use of deep learning techniques for optimizing channel sensing and downlink precoding for the massive MIMO systems and for systems equipped with intelligent reflecting surfaces. Instead of the traditional paradigm of first performing channel estimation, then designing the communication system according to the estimated channel, we advocate an approach of directly mapping the received pilots to the precoders through machine learning. For the frequency-division duplex massive MIMO system, we show that deep learning can be used for the efficient and distributed channel sensing, quantization, feedback and downlink multiuser precoding task with rate-limited feedback, in effect solving a distributed source coding problem. For the intelligent reflecting surface system, we show that a graph neural network can learn to find the optimal reflective coefficients and the optimal beamformers without explicit channel estimation. We show that the proposed deep learning approach can significantly reduce the required pilot length. We further make progress toward finding generalizable architectures and toward the interpretability of the resulting solutions.