emergenCITY und MAKI laden ein zur Distinguished Lecture mit Professor Gustavo de Veciana von der University of Texas, Austin / USA. In seinem Vortrag am 06. Juni 2024 um 16:15 Uhr geht es um Scalable and Resilient Networked Learning Systems“. Der Vortrag ist öffetlich:

  • Datum: Donnerstag, 06. Juni 2024, 16:15 Uhr
  • Referent: Gustavo de Veciana, University of Texas, Austin / USA
  • Titel: „Scalable and Resilient Networked Learning Systems“
  • Ort: S3 20, Raum 111 (Rundeturmstr. 10; 64283 Darmstadt) und online über Zoom
Über den Vortrag:

Many next-generation learning systems enabling applications in, e.g., healthcare, energy, banking, AR/VR design and car/robot navigation, will be privacy-driven, distributed and large-scale, resulting in substantially increased exposure to network congestion/failures. In this talk we describe work towards developing new, as well as expanding traditional, engineering principles for the design of scalable and resilient networked learning systems. We explore these challenges, by focusing on Federated Learning (FL) based systems aimed at facilitating the training of global models across large numbers of distributed clients with potentially heterogeneous data. The talk centers on four interrelated themes wherein we combine the development of theoretical underpinnings, architecture, applications and protocol design. We first study how to achieve resilience to uncertainty in FL systems subject to intermittent client availability and uplink resource constraints. We propose and explore a novel unbiased algorithm that dynamically learns a client availability-dependent sampling strategy which asymptotically minimizes the impact of client-sampling variance on the global model’s convergence. Secondly, since FL systems are exposed to, or indeed the cause of, congestion across a wide set of network resources we consider the use of lossy compression to reduce the size of exchanged files and associated delays, at the cost of adding noise to model updates. We propose a policy, which judiciously adapts clients’ compression to varying network congestion, and asymptotically minimizes FL wall clock training time. Finally, if time permits, we will briefly discuss an approach to exploit the aggregative character of FL client model updates via in-network update aggregation across an overlay Data Aggregation Networks (DANs). These are akin to Content Delivery Network (CDN) overlays which are a core element managing the cost and performance in current network infrastructure. The deployment and optimization of DANs have potential to overcome the fundamental bottleneck at the FL server, but requires some thought to address stochastic delays/losses due to congestion and its impact on FL convergence. Overall, this talk explores systematic approaches to the design of scalable and resilient distributed learning systems over congestible networks. The research exemplifies synergies amongst information, queueing and learning theory towards the design of algorithms and protocols to support networked learning.

Bio des Referenten:

Gustavo de Veciana received his Ph.D. in electrical engineering from the U.C. Berkeley in 1993. He is currently a Professor and Associate Chair of the Department of Electrical and Computer Engineering and recipient of the Cockrell Family Regents Chair in Engineering at U.T. Austin. He served as the Director and Associate Director of the Wireless Networking and Communications Group (WNCG) from 2003-2007. His research focuses on the design, analysis and control networks, information theory and applied probability. Current interests include: measurement, modeling and performance evaluation; wireless and sensor networks; architectures and algorithms to design reliable computing and networked systems. Dr. de Veciana is currently an editor at large for the IEEE/ACM Transactions on Networking. He was the recipient of an NSF CAREER Award 1996, and a co-recipient of 7 best paper awards including the 2021 IEEE Communication Society W. Bennett Prize. In 2009 he was designated IEEE Fellow for his contributions to the analysis and design of communication networks. He currently serves on the board of trustees of IMDEA Networks Madrid.