A round of federated training
A typical federated learning round follows a repeating cycle. The central server selects a subset of available devices or sites and sends them the current shared model. Each participant trains that model for a short time on its own local data, usually just a few passes, since the goal is a small local improvement, not a full training run from scratch.
Each participant then sends its updated model parameters, or the difference between the new and old parameters, back to the server. The server averages these updates together, weighting them in some reasonable way, such as by how much local data each participant used, and applies the result to produce a new, improved shared model. That new model becomes the starting point for the next round, and the cycle repeats, often for many rounds, until the model converges.