Making AI Sustainable: Online Optimization of Carbon and Energy in Cloud-Edge AI Systems

Add to your list(s) Download to your calendar using vCal

• Speaker to be confirmed
• Friday 23 May 2025, 15:00-16:00
• Computer Lab, FW11 and Online (MS Teams link to appear below).

If you have a question about this talk, please contact Richard Mortier.

Cloud-edge computing has become a critical enabler for realizing the potential of AI in the coming decade and beyond. Yet, AI systems across cloud-edge continuum often entail substantial energy consumption and significant carbon footprint. A promising approach to achieving carbon neutrality without compromising system efficiency is to leverage cap-and-trade programs, wherein a carbon allowance cap is obtained for a projected period and allowances can be then traded as needed in that period. In this talk, firstly, I will address the problem of carbon-neutral edge AI inference under such a framework. This setting poses several non-trivial challenges, including the unknown stochastic data distributions and arrival patterns, the exploration-exploitation trade-off under model switching costs, and the variability of allowance prices and system states. I will present our formulation of a long-term stochastic cost optimization problem that captures these challenges, alongside a learning-centric decomposition-based online algorithmic approach that adaptively samples models to minimize expected inference loss with bounded switching, while trading carbon allowances efficiently in real time without relying on future prices or emissions. I will also describe our theoretical guarantees and empirical validation of this approach. Subsequently, I will introduce our related work on energy-aware federated learning in cloud-edge environments, focusing on managing the energy usage of concurrent training jobs during demand response events while pursuing decarbonization. Finally, I will briefly highlight our additional efforts on energy-efficient diffusion-based generative AI and energy-constrained federated learning incentives, and conclude with a discussion of future research directions.

Biography: Lei Jiao received his Ph.D. in computer science from the University of Göttingen, Germany, in 2014. He is currently a faculty member at the University of Oregon, USA , and was previously a member of the technical staff at Nokia Bell Labs, Ireland. He researches networking and distributed computing, spanning AI infrastructures, cloud/edge networks, energy systems, cybersecurity, and multimedia. His work integrates mathematical methods from optimization, control theory, machine learning, and economics. He has authored over 80 peer-reviewed publications in journals such as IEEE Transactions on Networking, IEEE Transactions on Mobile Computing, IEEE Transactions on Parallel and Distributed Systems, and IEEE Journal on Selected Areas in Communications, and in conferences such as INFOCOM , MOBIHOC, ICDCS , SECON, ICNP , ICPP, and IPDPS , garnering over 6,000 citations according to Google Scholar. He is a recipient of the U.S. National Science Foundation CAREER Award, the Ripple Faculty Fellowship, the Alcatel-Lucent Bell Labs UK and Ireland Recognition Award, and several Best Paper Awards including those from IEEE CNS 2019 and IEEE LANMAN 2013 . He has served in various program committee roles, including as a track chair for ICDCS , as a member for INFOCOM , MOBIHOC, ICDCS , and WWW , and as a chair for multiple workshops with INFOCOM and ICDCS .

This talk is part of the Computer Laboratory Systems Research Group Seminar series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• CL's SRG seminar
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge talks
• Chris Davis' list
• Computer Lab, FW11 and Online (MS Teams link to appear below)
• Computer Laboratory Systems Research Group Seminar
• Department of Computer Science and Technology talks and seminars
• Interested Talks
• School of Technology
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.