Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
(LocalGradients, LocalControlVariate, GlobalControlVariate) -> CorrectedGradients
Correct local gradient step drift by tracking and applying the difference between local and global control variates.
Problem it solves
Heterogeneous client data distributions (non-I.I.D.) cause local models to drift away from the global optimum during local training epochs.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.