asynchronous federated learning

**Asynchronous Federated Learning** is a **federated learning approach where the server updates the global model immediately upon receiving any client's update** — without waiting for all selected clients to finish, eliminating the synchronization barrier that slows down FL with heterogeneous clients. **Asynchronous FL Approaches** - **FedAsync**: Server applies each client update immediately with a mixing coefficient. - **Staleness Weighting**: Weight client updates by their staleness ($alpha^{t - t_k}$) — old updates get less weight. - **Buffered**: Wait for a buffer of $K$ updates before aggregating — semi-synchronous middle ground. - **Federated Buffer**: Collect updates in a buffer and aggregate when buffer is full. **Why It Matters** - **No Stragglers**: Synchronous FL waits for the slowest client — async FL is not bottlenecked by stragglers. - **Throughput**: Higher model update frequency — more updates per unit time. - **Challenge**: Stale updates can degrade convergence — staleness mitigation is essential. **Async FL** is **don't wait, update now** — processing client updates as they arrive for continuous, straggler-free model improvement.