Objective Oriented Personalization in Federated Learning
We propose and analyze personalization models for various objectives (maximum participation, maximum welfare, fairness) in the Bayesian Hierarchical setting for Mean Estimation and Federated Learning. Further, we derive experimental results illustrating the performance of different personalization models with respect to these objective functions.
Demystifying the Price of System Heterogeneity in Federated Learning
In this work we explore the trade-off between data and system heterogeneity in Federated Learning. We propose and analyze a novel, asynchronous method that adaptively selects the aggregation weights at every round based on the quality of the model, the speed, and the level of the clients' heterogeneity. Our experimental results indicate that our proposed scheme enjoys significant benefits in terms of accuracy and convegence speed compared to traditional federated learning baselines.