Reproduce “MimiC: Combating Client Dropouts in Federated Learning by Mimicking Central Updates” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuchang Sun, Yuyi Mao, Jun Zhang
MimiC is a server-side algorithm developed by researchers at The Hong Kong University of Science and Technology and The Hong Kong Polytechnic University to combat client dropouts in federated learning. It modifies local updates using correction variables to mimic central updates, demonstrating theoretical convergence to a stationary point and achieving higher accuracy and faster convergence than baselines across diverse dropout scenarios.
Key insights
- FedAvg's convergence failure under arbitrary client dropouts is primarily due to a persistent 'objective bias' introduced by the aggregated update not reflecting the true global gradient.
- Although individual local updates can become stale, the *deviation* of each local update from the central update tends to remain relatively stable, enabling effective correction using historical information.
- Convergence can be theoretically guaranteed even with arbitrary dropouts by strategically modifying learning rates to ensure the gradient estimation error diminishes over time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.