Summary
In this chapter, we covered several FL frameworks through the context of two different examples. From the first example, you learned how a traditional centralized ML problem can be converted into the analogous FL scenario by separating the data into disjointed subsets. It is now clear that random sampling leads to local datasets that are IID, allowing FedAvg to reach the same level of performance as the centralized equivalent with any of the FL frameworks.
In the second example, you learned one of the many ways a group of datasets can be non-IID (different class label distributions) and observed how different severities of non-IID datasets affect the performance of FedAvg. We encourage you to explore how alternative aggregation methods can improve on FedAvg in these cases.
Both examples also should have given you a solid understanding of the general trends when working with different FL frameworks; while the specific implementation-level details may change (due to the...
