ATHENA INSTITUTE SEMINAR SERIES: Dynamic Decentralized Federated Learning

Presenter: Shenghong Dai, University of Wisconsin Madison

Abstract:

More vehicles are equipped with sensors that could collect massive data of wide-area environments. To leverage such data and protect clients' privacy, we naturally think of Federated Learning (FL) which trains a high-quality shared model through training decentralized data over clients and sending back only the model updates. However, a bottleneck might occur over the central server with a large amount of clients. This limitation motivates the need for decentralized FL where clients share their model updates with their neighbors instead of the central coordinator. We present a new decentralized FL algorithm with convergence guarantees to address two challenges: a) clients do not have static data but dynamically changing data; b) the connectivity graph is changing over time (e.g., nearby vehicles are not fixed). Finally, we develop a new decentralized FL simulator that provides a realistic modeling for the dynamic graph.

Bio:

Shenghong Dai is a Ph.D. candidate at the Department of Electrical and Computer Engineering at University of Wisconsin–Madison. Her research interests lie in the fields of machine learning, computer vision and mobile systems. In particular, her recent work focuses on distributed machine learning (e.g., federated learning) and on-device machine learning.