EECS Seminar: NextG Signal Processing Architectures - from mmWave to Deep Learning

Upamanyu Madhow, Ph.D.

ECE Department
UC Santa Barbara

Abstract: We illustrate in this talk the importance of a signal processing perspective for the design of NextG infrastructures by discussing a selection of problems from three application areas: communication, sensing and inference.

  1. Communication: We discuss recent work in hardware/signal processing co-design for realizing the immense potential of the mmWave/THz frequency bands for communication. Our goal is to scale “mostly digital” transceiver architectures while exploiting 10s of GHz of channel bandwidth and 100s of antennas. We introduce architectures and algorithms that seek to address hardware bottlenecks such as phase noise, nonlinearities and low-precision analog-to-digital conversion. Key concepts include spatial oversampling, exploiting the availability of a massive number of antenna elements and beamspace processing, exploiting the sparsity of the mmWave channel. 
  2. Sensing: We introduce a novel compressive MIMO radar architecture that is well matched to the concept of joint communication and sensing (JCAS). The architecture exploits large RF beamforming antenna array technologies developed for communication in a manner that increases range while maintaining a large field of view and enhancing spatial resolution. 
  3. Inference: Deep neural networks (DNNs) are pervasive in our cyberlives, but they are brittle black boxes that cannot be trusted in safety-critical applications. With standard end-to-end DNN training, we do not control or understand the features being extracted. We advocate supplementing end-to-end training with a signal processing approach explicitly aimed at shaping features at intermediate layers of the DNN, and report on encouraging preliminary results on enhanced robustness via feature shaping motivated by communication theory and neuroscience.

Bio: Upamanyu Madhow is Distinguished Professor of Electrical and Computer Engineering at UC Santa Barbara. His current research interests focus on next generation communication, sensing and inference infrastructures, with emphasis on millimeter wave systems, and on fundamentals and applications of robust machine learning. Madhow is a recipient of the 1996 NSF CAREER award, co-recipient of the 2012 IEEE Marconi prize paper award in wireless communications and recipient of a 2018 Distinguished Alumni award from the ECE department at the University of Illinois, Urbana-Champaign. He has served as associate editor for the IEEE Transactions on Communications, the IEEE Transactions on Information Theory and the IEEE Transactions on Information Forensics and Security. He is the author of two textbooks published by Cambridge University Press, Fundamentals of Digital Communication (2008) and Introduction to Communication Systems (2014). Madhow is a co-inventor on 31 U.S. patents and has been closely involved in technology transfer of his research through several start-up companies, including ShadowMaps, a software-only approach to GPS location improvement, which has been deployed worldwide by Uber.