When: Apr 12th, 2023 – 10:00 – 13:00 AM
Where: In presence at the San Niccolò building
Bayesian Inference and Generative Models for Wireless Communication Algorithms
Abstract:
This tutorial first introduces the fundamentals of generative models, followed by the perspective of learning the radio propagation environment (“radio-compatible digital twin”) of wireless communication scenarios, with the goal of providing prior information about the channel state information that is leveraged to enhance different types of physical layer functionalities in wireless communication systems. The tutorial begins with an introduction to Bayesian inference and generative models, showing that the recently proposed Variational Autoencoding (VAE) technique and the classical well-known Gaussian Mixture Models (GMM) method are just different versions of the same conceptual methodology. Both variants are excellent for modeling the distribution of channel state information (CSI) underlying a propagation environment as conditional normal distributions and allow the inference of corresponding prior information. In a further part of the tutorial, it will therefore be shown how this prior information can be used to estimate CSI as a first application example. It will be shown that and how the structural features of the statistical properties of the CSI, which are due to different access techniques and antenna technologies, can be exploited to further improve the estimation quality by appropriately adapting and modifying the standard architectures of the generative models used. In another chapter, we will present a Markov chain-based extension of standard VAEs, which is suitable for estimating time-varying channel state information, for example. For this purpose, the audience will first be introduced to a probability graph perspective on generative models. In the second application part, the tutorial will introduce the feedback generation of CSI supported by generative models, as well as the use of so-called adaptive codebooks, which are completely based on the exploitation of prior information with generative models.
Biography:
Wolfgang Utschick (Fellow, IEEE) received the diploma and doctoral degrees in electrical engineering with distinction from the Technical University of Munich (TUM), Germany. Prior to that, he completed several years of certified industrial training programs. Since 2002, he is a Professor with TUM for Methods of Signal Processing and since 2011, he is a TUM Asia faculty member in Singapore and a regular guest Professor with Singapore Institute of Technology. In 2021, he became a core member of the newly founded Munich Data Science Institute. From 2017 to 2022, Wolfgang Utschick served two terms of office as the dean of the Department of Electrical and Computer Engineering at TUM. In 2023, he has been appointed to a “Lighthouse
Professorship” at the new TUM School of Computation, Information and Technology. Wolfgang Utschick teaches courses on signal processing, stochastic processes, optimization theory, and machine learning in the field of wireless communications and other application areas of signal processing, he holds several patents in the field of multiantenna signal processing and has authored and coauthored a large number of technical articles in international journals and conference proceedings. He has edited several books and is the founder and the editor of the Springer book series Foundations in Signal Processing, Communications and Networking. Wolfgang Utschick has been awarded several best paper prizes and has been involved as principal investigator in several research projects funded by the German Research Fund (DFG) and coordinated a German DFG priority program on Communications Over Interference Limited Networks. He is a Fellow of the IEEE and Chair-elect
of the VDE German Information Technology Society’s expert group KT1 on Information Theory and System Theory.