Understanding FedAvg

In Chapter 3, Workings of the Federated Learning System, the aggregation algorithm known as FedAvg was introduced to help clarify the general structure and represent the more abstract concepts discussed earlier with a specific example. FedAvg was used for two reasons: simplicity in the underlying algorithm, and generalizability across more model types than gradient-based approaches. It also benefits from extensive references by researchers, with performance analysis in different theoretical scenarios using FedAvg as a baseline when proposing new aggregation methods. This focus in the research community can most likely be attributed to the fact that the original FedAvg paper was published by the team working at Google that first brought exposure to the concept and benefits of FL. For further reading, this paper can be found at https://arxiv.org/abs/1602.05629?context=cs.

FedAvg is predated by an aggregation approach known as Federated Stochastic Gradient Descent...