Date: Mardi 16 Juin 2026, 11h00
Lieu: salle C057
Title: FedStaleAsync: Efficiently Reusing Historical Updates to Enhance Asynchronous Federated Learning
_________________________________________________________
Résumé : Federated learning (FL) enables clients to collaboratively train a shared model without exchanging local data. While asynchronous federated learning (AFL) improves training efficiency under high device heterogeneity by removing the synchronization bottlenecks of standard FL, it introduces stale updates and may exacerbate participation imbalance among heterogeneous clients. To mitigate this issue, prior work has proposed reusing historical updates to increase the contribution of slower clients. We theoretically show that existing AFL methods relying on this mechanism tend to underutilize newly received updates, placing excessive emphasis on increasingly stale information. Although strictly unbiased aggregation can correct this issue, it often incurs a substantial increase in variance. To address this trade-off, we propose FedStaleAsync, a principled weakly unbiased aggregation method that better exploits newly transmitted updates without significantly increasing variance. Experiments across diverse heterogeneous settings demonstrate that FedStaleAsync consistently outperforms existing asynchronous methods, achieving faster convergence and higher final accuracy.
_________________________________________________________