A new way to update models for moving devices
Researchers developed F-SOSA to help models adapt quickly when devices join or leave a network.
Researchers created F-SOSA, a stable one-shot federated learning algorithm. This method allows for rapid and exact model editing.
The system handles device dynamics like new client contributions or the withdrawal of previous contributions. It uses F-SOSA-U for unlearning and F-SOSA-I for incremental learning.
One-shot federated learning (FL) completes training in a single communication round. Model editing removes withdrawn contributions or incorporates new ones without retraining from scratch.
F-SOSA uses sub-sampling of local models with a carefully tuned rate for stable aggregation. It also performs curvature-aware model fusion to preserve utility.
The edited model and its states are statistically indistinguishable from retraining. This process incurs only an expected fraction of the cost, where quantifies stability.
In urban mobility simulations, this approach matches retraining performance. It reduces cumulative time by at least 36% and communication overhead by at least 22%.
Key facts
- F-SOSA reduces cumulative time by at least 36% in urban mobility simulations.
- The method reduces communication overhead by at least 22%.
- F-SOSA-U and F-SOSA-I enable unlearning and incremental learning in a single round.