In their newest paper on Implicit Model Specialization through DAG-based Decentralized Federated Learning, Jossekin Beilharz, Bjarne Pfitzner (from the Connected Healthcare Group), Robert Schmid, Paul Geppert together with Bert Arnrich and Andreas Polze show that specialized machine learning models can evolve from a fully decentralized federated learning consensus in a learning DAG. Moreover, concerns such as weighing specialization vs. generalization and poisoning robustness in the decentralized learning scheme are addressed.
The authors will present their work in the Systems Support for Machine Learning session on Thursday, December 9 at 17:15 CET. Registration for ACM Middleware 2021 is free. The preprint is available now on ArXiv.