Data-Intensive Distributed Systems LaboratoryCS Seminar
Date: April 1st, 2026
Time: 12:50pm
Room: SB111
Dr. Nathaniel Hudson
Assistant Professor
Computer Science Department
Illinois Institute of Technology
Talk Title
Learning from resource-constrained devices with model-heterogeneous federated learning
Talk Abstract
Most data generated each day come from decentralized sources (e.g., smartphones, wearables, sensors). These valuable data contain knowledge that can be used to train state-of-the-art AI models. However, the conventional approach to training state-of-the-art AI requires mass collection of these data to train AI on centralized compute infrastructure (e.g., data centers, cloud). This approach is unappealing for several reasons. Chief among them, concerns related to data privacy and concerns regarding data transfer costs come to mind.
One recent paradigm to side-step this issue is to train AI models where the data naturally live: Federated Learning (FL). With FL, instead of transferring the data from the sources to where the AI model is trained, we instead send the AI model to where the data naturally live. This approach improves communication efficiency and data privacy from conventional approaches which rely on centralized model training. However, the limited resources available on low-power edge devices can be prohibitive. A low-power sensor might have valuable data but insufficient hardware resources to train a complex AI model.
A recent flavor of FL has since been proposed to address this issue. Model-heterogeneous FL adopts model pruning methods to send AI models of different sizes to devices based on their hardware constraints to create more inclusive FL systems. In this talk, we will discuss the state-of-the-art in FL and model-heterogeneous FL and some of the open problems that I am actively exploring now in my own research.
Speaker Bio
Nathaniel Hudson is an Assistant Professor of Computer Science at the Illinois Institute of Technology, where he works on decentralized machine learning systems. His research focuses on federated and decentralized learning, cyber-physical systems, AI for science, and complex networks, with an emphasis on serving AI on edge computing infrastructure for smart city applications. His recent work includes systems for hierarchical federated learning, QoS-aware edge AI placement and scheduling, and federated reinforcement learning for traffic control.