Experience-driven network topology construction for efficient decentralized federated learning

Decentralized federated learning (DFL) has evolved as a favored paradigm for cultivating machine learning models on extensive data in edge computing, thanks to its prowess in circumventing potential bottlenecks inherent in conventional parameter server architectures. However, existing DFL solutions predominantly leverage deterministic topologies, contending with system heterogeneity and non-independent and identically distributed (non-IID) local datasets. This dilemma often escalates bandwidth costs and extends convergence rates within fluctuating networks. To this end, we present a groundbreaking mechanism named data-efficient decentralized federated learning (DE-DFL), specifically designed to accelerate the model training process. In DE-DFL, each client interacts with its neighbors, e.g., model exchange, according to an approximate policy at every round, so as to reduce bandwidth consumption. Subsequently, we then propose an experience-driven algorithm to adaptively determine the optimal communication policy for all clients according to real-time system situations (e.g., data distribution and bandwidth resource). Our innovative mechanism has been rigorously validated against standard models and datasets, thereby corroborating its superior efficacy. The experimental results reveal that DE-DFL significantly reduces the model training completion time by approximately 68.7% and enhances test accuracy by 6.9% under bandwidth constraints when compared to existing state-of-the-art solutions.