Initialisation and network effects in decentralised federated learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Network Science

Background:

• Decentralized federated learning (DFL) enables collaborative model training without central coordination, preserving data privacy.
• DFL's performance is sensitive to network topology and initial model conditions.
• Existing DFL methods often lack efficient, uncoordinated initialization strategies.

Purpose of the Study:

• To propose a novel initialization strategy for DFL that leverages network topology.
• To improve the training efficiency and scalability of decentralized artificial neural networks.
• To investigate the impact of network structure on DFL dynamics.

Main Methods:

• Developed an uncoordinated initialization strategy for artificial neural networks based on eigenvector centrality distribution.
• Analyzed the influence of network topology on DFL performance.
• Studied the scaling behavior and parameter choices under the proposed initialization.

Main Results:

• The proposed initialization strategy significantly enhances decentralized federated learning efficiency.
• Demonstrated a clear link between network topology, initialization, and learning dynamics.
• Identified optimal environmental parameters for the proposed initialization strategy.

Conclusions:

• Network structure and initialization are critical for efficient DFL.
• The proposed method offers a scalable and effective approach for decentralized AI training.
• This research provides foundational insights for designing robust DFL systems.