Reproduce “FedNSAM:Consistency of Local and Global Flatness for Federated Learning” 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
Junkang Liu, Fanhua Shang, Yuxuan Tian, Hongying Liu, Yuanyuan Liu
Researchers from Tianjin University and Xidian University introduce FedNSAM, an algorithm designed to harmonize local and global flatness in federated learning environments, particularly under data heterogeneity. It achieves a tighter theoretical convergence bound and a reduced 'flatness distance' compared to FedSAM, empirically demonstrating superior generalization (e.g., 66.04% accuracy for ResNet-18 on CIFAR100) and significantly faster convergence (e.g., 3x acceleration) across various CNNs and Vision Transformers.
Key insights
- The 'flatness distance' provides a measurable quantity that correlates with data heterogeneity and the inconsistency of local vs. global flatness, highlighting why local SAM efforts may not translate to global generalization.
- Leveraging global momentum for both perturbation direction and Nesterov extrapolation allows clients to consistently align their search for flat minima, leading to a flatter global loss landscape.
- FedNSAM achieves a tighter theoretical convergence bound for non-convex functions and a superior upper bound for the flatness distance compared to FedSAM, mitigating the negative impact of data heterogeneity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.