Henry Su Wang - Communication systems
Date: - Location: Eurecom
Abstract: Federated learning (FL) enables the collaborative development of machine learning (ML) models for modern networks. The integration of FL and edge/fog networks poses several challenges, however. The very features that make edge/fog systems so flexible - data and structural heterogeneity, flexible topologies, resource constraints, etc. - also make FL optimization and convergence challenging. In this talk, I explore network-aware FL, which explicitly leverages underlying network structure and heterogeneity to improve optimization, scalability, and convergence. I aim to develop methodologies to systematically exploit and define forms of edge/fog network heterogeneity in order to better understand core principles of FL over networks. The talk first examines device/client selection for aggregations, and then extends FL to emerging forms of edge/fog network heterogeneity, including labeled/unlabeled data imbalance and heterogeneous data features and/or modalities. Short bio: Henry Su Wang is a joint Lecturer and Postdoctoral Research Associate in Electrical and Computer Engineering at Princeton University. His research lies at the intersection of federated learning, distributed optimization, and networked systems, with a focus on their application on heterogeneous edge and fog networks. He received his Ph.D. and B.S.E.E. (with distinction) from Purdue University.
