Model Aggregation

In the Model aggregation basics section of Chapter 3, Workings of the Federated Learning System, we introduced the concept of aggregation within the federated learning (FL) process at a high level. Recall that aggregation is the means by which an FL approach uses the models trained locally by each agent to produce a model with strong global performance. It is clear to see that the strength and robustness of the aggregation method employed are directly correlated to the resulting performance of the end global model.

As a result, choosing the appropriate aggregation method based on the local datasets, agents, and FL system hierarchy is key to achieving good performance with FL. In fact, the focal point of many publications in this field is providing mathematically backed convergence guarantees for these methods in a variety of theoretical scenarios.

The goal of this chapter is to cover some of the research that has been done on aggregation methods and their convergence...