Core Concepts
The proposed FedACG algorithm improves the consistency across clients and facilitates the convergence of the server model by broadcasting a global model with a lookahead gradient, enabling clients to perform local updates along the trajectory of the global gradient. FedACG also regularizes local updates by aligning each client with the overshot global model to reduce bias and improve the stability of the algorithm.
Abstract
The paper proposes a novel federated learning algorithm called Federated Averaging with Accelerated Client Gradient (FedACG) to address the challenges of high heterogeneity in training data distributed over clients and limited client participation rates in federated learning.
Key highlights:
FedACG transmits the global model integrated with the global momentum as a single message, allowing each client to perform local updates along the landscape of the global loss function. This approach reduces the gap between global and local losses.
FedACG adds a regularization term in the objective function of clients to make the local gradients more consistent across clients, further improving the stability of the algorithm.
FedACG is free from additional communication costs, extra computation in the server, and memory overhead of clients, making it suitable for real-world federated learning settings.
FedACG demonstrates outstanding performance in terms of communication efficiency and robustness to client heterogeneity, especially with low client participation rates, outperforming state-of-the-art federated learning methods.
The authors provide a theoretical convergence analysis of FedACG for non-convex loss functions, matching the best convergence rate of existing federated learning methods.
Stats
The paper does not contain any explicit numerical data or statistics to support the key logics. The results are presented in the form of accuracy and communication rounds.