Accelerating Federated Learning for Edge Intelligence Using Conjugated Central Acceleration With Inexact Global Line Search

Driven by the increasing demand for real-time, low-latency learning processes and the ever-growing emphasis on data privacy, Federated Learning (FL) enabled edge intelligence emerges as a promising decentralized learning paradigm at the edge of the network, which empowers collaborative model training on edge agents, allowing them to make intelligent decisions locally without relying solely on centralized cloud servers. To enhance the training efficiency of edge agents and alleviate communication burdens, we propose a novel technique called Conjugated Central Acceleration with Inexact Line Search enabled Federated Stochastic Variance Reduced Gradient (CLSFSVRG). Conjugate Central Acceleration leverages conjugate gradient technique to efficiently utilize the training information from multiple edge agents by additional updating efforts in the central server, thereby enhancing the convergence rates of the global model and reduce the local training burden. Inexact Line Search optimizes the step size for model updates, striking a balance between precision and computational efficiency. Simulation results demonstrate that the proposed scheme outperforms the state-of-the-art FL algorithms, achieving superior performance in terms of higher test accuracy and faster convergence speed. Remarkably, our approach reduces communication costs by an impressive 82.4%, while still achieving a test accuracy of 96.5%. By allowing a small portion of edge agents to participate, CLSFSVRG exhibits higher robustness without compromising the test accuracy. Moreover, the fast convergence speed achieved with a limited number of participating edge agents contributes to significant reductions in edge computing cost during the training procedure.