Core Concepts
FedFisher is a novel algorithm for one-shot federated learning that leverages Fisher information matrices to improve communication and computation efficiency.
Abstract
The content introduces FedFisher, a one-shot FL algorithm addressing drawbacks of standard FL methods. It discusses theoretical analysis, practical implementations using Fisher information, and extensive experiments showcasing improved performance over baselines.
Introduction to decentralized data collection and FL framework.
Challenges in standard FL due to data heterogeneity across clients.
Proposal of FedFisher algorithm leveraging Fisher information matrices for one-shot FL.
Theoretical analysis for two-layer over-parameterized neural networks.
Practical implementation variants using diagonal Fisher and K-FAC approximations.
Extensive experiments demonstrating consistent improvements over competing baselines on various datasets.
Stats
"Extensive experiments on various datasets show consistent improvement over competing baselines."
"FedFisher variants consistently outperform other baselines by almost 10-20%."
"For CIFAR10 dataset, FedFisher(K-FAC) achieved an accuracy of 80.42%."
"Empirical validation shows that the error decreases as the width of the model increases."