Federated Averaging (FedAvg) is a core algorithm in federated learning that enables collaborative model training across decentralized devices or servers while keeping data localized. Its primary role in optimization is to coordinate updates from multiple participants—such as mobile devices or edge nodes—to iteratively improve a shared machine learning model without transferring raw data. Instead of centralizing data, FedAvg aggregates parameter updates (e.g., gradients or weights) computed locally by each participant, averages them, and applies the result to a global model. This approach addresses challenges like data privacy, network bandwidth constraints, and heterogeneous data distributions.
The algorithm works in repeated rounds. First, a central server sends the current global model to a subset of participants. Each participant trains the model locally using their own data, computes updates (e.g., adjusting neural network weights via stochastic gradient descent), and sends these updates back to the server. The server then averages these updates—often using a simple arithmetic mean—to produce a new global model. For example, in a mobile keyboard prediction scenario, thousands of devices might train a language model locally on user typing data. FedAvg averages these updates to create a global model that reflects diverse usage patterns without exposing individual keystrokes. To handle uneven data or device availability, techniques like weighted averaging (assigning higher weight to devices with more data) or partial participant sampling are often used.
FedAvg’s optimization benefits stem from balancing efficiency and robustness. By performing local training, it reduces communication overhead—devices can run multiple training steps before sending updates, minimizing server-client interactions. It also tolerates non-IID (non-independent and identically distributed) data, as participants’ local datasets may vary widely. For instance, in healthcare, hospitals with different patient demographics can collaboratively train a diagnostic model without sharing sensitive records. However, FedAvg introduces challenges like potential staleness (slow or dropped participants) and convergence instability. Developers often address these by tuning hyperparameters (e.g., learning rates, local epochs) or using adaptive optimization techniques (e.g., server-side momentum). Frameworks like TensorFlow Federated and PyTorch Lightning implement FedAvg as a baseline, allowing developers to adapt it for specific use cases while maintaining data privacy and scalability.