NeurIPS 2020

Distributionally Robust Federated Averaging

Meta Review

All reviewers agree (and I concur) that the paper contains solid theoretical results that improve in a non-trivial way upon previous work in FL. The overall impression was that the paper should be accepted based on its theoretical contributions. There were however some serious concerns regarding the experimental setup (that allows for trivial baselines that outperforms the evaluated methods); I urge the authors to seriously reconsider their methodology and ideally make a complete overhaul to the experimental part of the paper. (When doing so, please inspect carefully the comments of R1 for detailed guidelines.)