Federated Learning on the Edge

Federated learning reverses the training direction: instead of collecting data to a central model, the model travels — the server ships weights to devices, each trains on its local data, devices send back only updates (gradients/weight deltas), and the server aggregates. FedAvg, the canonical algorithm, is just a weighted average of client updates — each weighted by its data size.

The canonical deployment is Google Gboard's next-word prediction: trained on millions of phones, on the typing data that never leaves them. The privacy promise: raw data stays on-device; only model updates travel.