Abstract: The emergence of networks constituted by a massive number of devices with sensing, computation, and wireless capabilities presents new opportunities for large-scale machine learning (ML). Yet, leveraging the enormous amount of data generated by these devices brings new challenges, associated with the decentralized nature of the system and resource-limited wireless connectivity. In this talk, we focus on fully-decentralized learning architectures, where a number of devices, such as a swarm of unmanned aerial vehicles operating in a remote area, aim to solve a decentralized learning problem. Decentralized gradient descent (DGD) is a renowned algorithm to achieve this goal: it relies on iterative consensus (communications) and local gradient descent (computations) updates over a predetermined graph. Yet, a naïve implementation of DGD over wireless channels requires topology information on the connectivity structure of the network (who communicates to whom), power/rate control algorithms to combat fading, and careful interference management (e.g., via TDMA scheduling). These practical aspects make DGD not scalable over wireless networks with massive number of devices. In this talk, I will propose a scheme that allows one-shot consensus (i.e., all devices transmit simultaneously, in contrast to a TDMA-based approach), by leveraging the waveform superposition properties of the wireless channels. The key mechanism is a non-coherent over-the-air physical layer scheme, representing a noisy-version of the consensus step of DGD. I will show that, in contrast to naïve implementations of DGD, the proposed scheme operates without channel state information and leverages the channel pathloss to mix signals, without explicit knowledge of the graph Laplacian matrix, nor of the graph topology. I will finally demonstrate, both theoretically and numerically, that, with a suitable tuning of consensus and learning stepsizes, the error (measured as Euclidean distance) between the local and globally optimum models vanishes with rate O(k−1/4) after k iterations. 
 
Bio: Nicolò Michelusi (Senior Member, IEEE) received the B.Sc. (with honors), M.Sc. (with honors), and Ph.D. degrees from the University of Padova, Italy, in 2006, 2009, and 2013, respectively, and the M.Sc. degree in telecommunications engineering from the Technical University of Denmark, Denmark, in 2009, as part of the T.I.M.E. double degree program. From 2013 to 2015, he was a Postdoctoral Research Fellow with the Ming-Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, CA, USA, and from 2016 to 2020, he was an Assistant Professor with the School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, USA. He is currently an Associate Professor with the School of Electrical, Computer and Energy Engineering, Arizona State University, Tempe, AZ, USA. His research interests include 5G wireless networks, millimeter-wave communications, stochastic optimization, machine-learning over wireless systems. He served as Associate Editor for the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS from 2016 to 2021, and currently serves as Editor for the IEEE TRANSACTIONS ON COMMUNICATIONS. He was the Co-Chair for the Distributed Machine Learning and Fog Network workshop at the IEEE INFOCOM 2021, the Wireless Communications Symposium at the IEEE Globecom 2020, the IoT, M2M, Sensor Networks, and Ad-Hoc Networking track at the IEEE VTC 2020, and the Cognitive Computing and Networking symposium at the ICNC 2018. He was the Technical Area Chair for the Communication Systems track at Asilomar 2023. He received the NSF CAREER award in 2021 and the IEEE Communication Theory Technical Committee (CTTC) Early Achievement Award in 2022.