Tailored Federated Learning With Adaptive Central Acceleration on Diversified Global Models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) faces challenges in aggregating diverse local models for individual machine benefit.
• A single global model may not align with each machine's unique interests and profit maximization goals.

Purpose of the Study:

• To develop accelerated federated training procedures that allow machines to leverage shared knowledge while achieving customized solutions.
• To address heterogeneous demands in FL by proposing novel grouping and acceleration mechanisms.

Main Methods:

• Introduced model-based grouping mechanism with adaptive central acceleration (MA-FSVRG) based on the federated stochastic variance reduced gradient (FSVRG) framework.
• Proposed gradients-based grouping mechanism with adaptive central acceleration (GA-FSVRG) within the FSVRG framework.
• Evaluated performance against state-of-the-art FL baselines through simulations.

Main Results:

• Both MA-FSVRG and GA-FSVRG demonstrate advantages over existing FL methods.
• MA-FSVRG offers enhanced performance stability and reduced local computation costs.
• GA-FSVRG achieves superior test accuracy and faster convergence, especially with limited machine participation.

Conclusions:

• The proposed MA-FSVRG and GA-FSVRG effectively tackle heterogeneous demands in federated learning.
• These methods enable personalized model optimization and improved collaborative efficiency in FL settings.