Federated learning with workload-aware client scheduling in heterogeneous systems.

Area of Science:

• Distributed Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Federated Learning (FL) enables local model training on edge devices without data centralization.
• Systems heterogeneity in FL leads to stragglers and reduced model accuracy.
• Resource-constrained edge devices exacerbate these challenges.

Purpose of the Study:

• To propose an adaptive federated framework to address systems heterogeneity in FL.
• To reduce stragglers and improve the robustness and convergence speed of global models.
• To enhance fairness in model performance across heterogeneous devices.

Main Methods:

• FedSAE: Leverages client workload completion history for adaptive workload prediction, reducing stragglers.
• Active Learning integration: Dynamically schedules participants based on training loss, accelerating convergence.
• q-FedSAE: Combines FedSAE with q-FFL for improved global fairness in heterogeneous systems.

Main Results:

• FedSAE significantly reduces stragglers (by 90.69% on average) and improves testing accuracy (by 22.19%) compared to FedAvg.
• Both FedSAE and q-FedSAE demonstrate faster convergence than FedAvg in heterogeneous environments.
• q-FedSAE achieves robust convergence and fairer model performance than q-FedAvg while maintaining similar accuracy to FedSAE.

Conclusions:

• FedSAE effectively mitigates issues arising from systems heterogeneity in Federated Learning.
• The proposed adaptive framework enhances model accuracy, reduces delays, and accelerates convergence.
• q-FedSAE offers a solution for achieving fairness alongside performance in heterogeneous FL settings.