T2: Communication Networks Design: Model-Based, Data-Driven, or Both?

Presented by: Alessio Zappone (CentraleSupelec), Marco Di Renzo (Paris-Saclay Univ.), Merouane Debbah (Huawei R&D)

Time: 14:00–17:30
Room: Milo 1

Abstract—The tutorial will provide the audience with a solid understanding of the fundamentals of deep learning and its use for the design of wireless communications. Artificial neural networks, which are the distinctive feature of deep learning as compared to other machine learning methods, will be introduced. The main artificial neural networks architectures will be described, focusing in particular on feedforward networks and of the problem of their supervised training. The most widely used methods for neural network training will be described and tips and tricks to improve the training process will be explained.

After introducing the fundamentals of deep learning, the tutorial will address how deep learning can be merged with more traditional model-based approaches to perform wireless networks design, exploiting the frameworks of transfer learning and reinforcement learning. Several relevant applications will be described, quantifying the advantages of embedding expert knowledge coming from theoretical models, into data-driven methods, considering diverse system scenarios, such as dense heterogeneous cellular networks, energy-efficient networks, network-slicing systems, and chemical-based communication systems. A main conclusion drawn by the tutorial is the embedding expert knowledge into traditional neural network design can significantly reduce the amount of data that is necessary to use for training purposes, thus significantly simplifying the overall system design.

Tutorial Objectives
Data-driven approaches are not new to wireless communications, but their implementation through deep learning techniques has never been considered in the past, even though deep learning is the most widely used machine learning approach in other fields than wireless communication. In our opinion, this is mainly due to the fact that, unlike other fields of science where theoretical modeling is particularly hard, thus motivating the use of data-driven approaches, wireless communications could always rely on strong mathematical models for system design. However, the situation is rapidly changing, and very recently the use of deep learning has started being envisioned for wireless communications too. Indeed, the increasing complexity of wireless networks makes it harder and harder to come up with theoretical models that are at the same time accurate and tractable. The rising complexity of 5G and beyond 5G networks is exceeding the modeling and optimization possibilities of standard mathematical tools.

Nevertheless, purely data-driven approaches require a huge amount of data to operate, which might be difficult and/or expensive to acquire in practical large-scale scenarios. This issue has been acknowledged also by the deep learning community, for which one major research trend lies precisely in the development of techniques able to exploit any prior information that is available about the problem at hand in order to reduce the amount of data needed to achieve given performance levels. In this context, the specific field of communication theory presents a major opportunity thanks to the availability of many more theoretical models compared to other fields of science. Indeed, despite being usually inaccurate and/or cumbersome, available communication models still provide important prior information that should be exploited. Accordingly, the overall aim of this tutorial is to put forth the idea that theoretical modeling and data-driven approaches are not two contrasting paradigms, but should rather be used jointly to get the most out of them.

This tutorial will cover the most recent approaches to merge advanced deep learning techniques with the latest model-based methodologies for system-level design and optimization of wireless networks. Specifically, the following objectives are pursued:

1) Provide the foundations of deep learning by artificial neural networks and introduce the main concepts about supervised training of neural networks.
2) Show how deep learning can complement, rather than replace, traditional wireless networks design methodologies, to develop novel design approaches with reduced complexity and improved performance.
3) Present a wide range of applications to evidence the gain that a joint data-driven / model-based approach can bring, as compared to using only one of the two approaches.

Tutorial Outline
Fundamentals of machine learning for communications A definition of machine learning Supervised vs. unsupervised learning Underfitting, overfitting, and capacity Deep learning vs. Machine learning Artificial neural networks Feedforward neural networks Training artificial neural networks Embedding models into deep learning Model-based or data-driven? Expert knowledge into artificial neural networks Learning to optimize Transfer learning Deep reinforcement learning Applications Resource allocation in dense heterogeneous cellular networks Resource allocation in network-slicing systems Resource allocation in energy-efficient cellular networks Optimization of beyond-RF wireless networks Concluding remarks

Primary Audience
The tutorial is aimed at both academic researchers wishing to learn the fundamentals of this emerging field, as well as to wireless engineers and industry practitioners wishing to employ deep learning to improve their products.

Novelty
This tutorial is the first to:
1) focus specifically on deep learning considering the cross-fertilization between data-driven methods and model-based approaches.
2) discuss emerging machine learning tools like deep transfer learning and deep reinforcement learning
3) Focus on the use of deep learning for the resource management of wireless networks, presenting many relevant applications for several instances of communication systems.

Biography
Dr. A. Zappone is currently an experienced Marie Curie Fellow at CentraleSupelec, France, working in the field of resource allocation for 5G wireless networks and beyond. He is an IEEE Senior Member, an Associate Editor of the IEEE Signal Processing Letters, and has been a Guest Editor of the IEEE JSAC Special issue on “Energy-Efficient Techniques for 5G Wireless Communication Systems”. He was appointed exemplary reviewer for both the IEEE Transactions on Communications and IEEE Transactions on Wireless Communications in 2017.

Dr. M. Di Renzo is Associate Professor with the Laboratory of Signals and Systems of Paris-Saclay University – CNRS, CentraleSupelec, Univ Paris Sud, France. He is a Distinguished Visiting Fellow of the Royal Academy of Engineering (UK), and co-founder of the university spin-off company WEST Aquila s.r.l., Italy. He serves as associate Editor-in-chief of the IEEE Communication Letters, and editor of the IEEE Transactions on Communications, (Heterogeneous Networks Modeling and Analysis) and Transactions on Wireless Communications. He is an IEEE Senior Member, an EURACON Member, and a Distinguished Lecturer of the IEEE Communications and IEEE Vehicular Technology Societies.

Dr. M. Debbah is Vice-President of the Huawei France R&D center and director of the Mathematical and Algorithmic Sciences Lab, as well as full professor at CentraleSepelec. He is an IEEE Fellow, a WWRF Fellow and a member of the academic senate of Paris-Saclay. He received several awards, among which the 2015 IEEE Communications Society Leonard G. Abraham Prize, the 2015 IEEE Communications Society Fred W. Ellersick Prize, the 2016 IEEE Communications Society Best Tutorial paper award, the 2016 European Wireless Best Paper Award, the 2017 Eurasip Best Paper Award and the 2018 IEEE Marconi Prize Paper Award, the Mario Boella award in 2005, the IEEE Glavieux Prize Award in 2011, and the Qualcomm Innovation Prize Award in 2012.

All speakers have extensive tutorial experience, being regular tutorial/keynote speakers.