EF-Feddr: communication-efficient federated learning with Douglas-Rachford splitting and error feedback

Area of Science:

• Distributed Machine Learning
• Optimization Algorithms

Background:

• Federated learning (FL) faces communication bottlenecks impacting efficiency.
• Non-IID data and frequent client-server communication hinder FL scalability.

Purpose of the Study:

• To propose EF-Feddr, a communication-efficient algorithm for federated composite optimization.
• To address challenges in FL with non-IID data and biased compression.

Main Methods:

• Developed EF-Feddr based on relaxed Douglas-Rachford splitting.
• Integrated error feedback (EF) to ensure convergence with biased compression.

Main Results:

• EF-Feddr achieves O(1/K) convergence rate and O(1/ε^2) communication complexity.
• Experiments on FEMNIST, Shakespeare, and synthetic data validate EF-Feddr's efficacy.

Conclusions:

• EF-Feddr effectively overcomes convergence degradation from biased compression.
• The algorithm provides a practical solution for communication-constrained federated learning.