Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
(List<ClientModel>, ProxyDataset) -> DistilledGlobalModel
Train a global server model by distilling the ensemble predictions of heterogeneous client models on a shared proxy dataset.
Problem it solves
Direct parameter averaging (FedAvg) fails when client model architectures are diverse or highly heterogeneous.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.