Tutorial #4 – Low Resolution Signal Processing in Communications Using a Machine-Learning Framework

Instructor: Jan Lewandowsky, Hamburg University of Technology, Germany
Instructor: Maximilian Stark, Hamburg University of Technology, Germany

Abstract: The design of quantized algorithms for communication systems using mutual information as a design criterion has recently attracted considerable interest in the communications community. The fundamental idea is building quantized signal processing chains that aim to preserve the maximum possible relevant information in all involved signal processing algorithms, while using as few bits as possible for the signal representation and processing. This idea is different to conventional design approaches for quantized systems, which typically focus on minimizing a distortion measure for a given precision, for example, the mean squared error. A framework from machine-learning termed the Information Bottleneck method can be applied as a very powerful tool to design quantized signal processing algorithms with a focus on the preservation of relevant information. While this method is well known in the machine learning community, it is still rather unknown in the communications community. However, recent works describe very successful applications of the Information Bottleneck method in quantizer design, the design of low-density parity-check decoders for binary and nonbinary codes, the construction of polar decoders and detection schemes, as well as in massive MIMO systems. Most importantly, the Information Bottleneck approach allows obtaining coarsely quantized signal processing algorithms with very simple signal processing operations, but close-to-optimum performance. This tutorial covers the information theoretical ideas behind the Information Bottleneck method and explains how its information theoretical concept can be applied to build quantized signal processing algorithms for the aforementioned applications in detail. Moreover, it provides very intuitive and easy-to-understand examples that illustrate and visualize the presented information theoretical ideas in practice, to enable an easy understanding.

Bio: Jan Lewandowsky is a scientific researcher with the software defined radio group of the Fraunhofer Institute for Communication, Information Processing and Ergonomics (FKIE) in Wachtberg, Germany. He is also a Ph. D. candidate with the Institute of Communications at the Hamburg University of Technology (TUHH) in Hamburg, Germany. He has submitted his Ph.D. dissertation on signal processing approaches for communications based on the Information Bottleneck method to the dissertation committee of the TUHH in December 2019 and aims to graduate in 2020. From 2006 until 2018 he was an active officer with the German Air Force with a focus on radio communications and radar data processing. Jan Lewandowsky has received the B.Sc. degree and the M.Sc. degree (with distinction) in electrical engineering from the University of the Federal Armed Forces Munich in Neubiberg, Germany in 2010 and 2011, respectively. He received the first price award for his master thesis on robust communication in fast fading environments from the German chapter of Armed Forces Communications and Electronics Association in 2012. He has authored many journal and conference papers on the topic of this tutorial proposal and related topics. His research interests are mainly channel coding and modulation, as well as practical applications of information theory in communication systems. Jan Lewandowsky is 33 years old and lives in the Mid-Rhine region in Germany.

Bio: Maximilian Stark is a scientific researcher with the Institute of Communications at the Hamburg University of Technology (TUHH) in Hamburg, Germany. After having received the B.Sc. degree and the M.Sc. degree (with distinction) in electrical engineering from the TUHH in 2014 and 2017, respectively, he started to work on his Ph.D. on machine learning and signal processing with the Information Bottleneck method and related information theoretical concepts. In 2019, he joined the machine-learning group at the Nokia Bell Labs in Paris as a visiting researcher focusing on deep learning for communications. In the scope of his Ph.D. thesis, he also works in research cooperations with the Japan Advanced Institute of Science and Technology (JAIST), University of California (UCLA) and other cooperation partners. Already his master thesis focused on signal processing with the Information Bottleneck method. He has won the Karl H. Ditze award for his Master thesis in 2018. Channel coding and modulation also dominate his current research interests. Matters of particular interest in this context are machine-learning approaches for constellation design, detection and channel decoding as well as compressed sensing and massive-machine-type communication. Despite his young age, Maximilian Stark has authored and co-authored many papers in major technical journals and conferences already. Maximilian Stark is 27 years old and lives in Hamburg, Germany.