• Introduces SD‑GT, a semi‑decentralized gradient‑tracking method for fog‑based federated learning. • Eliminates gradient‑diversity assumptions, enabling robust training across highly heterogeneous subnetworks. • Provides convergence bounds for non‑convex, convex, and strongly‑convex objectives analysis. • Optimizes trade‑off between communication cost and model quality via subnet sampling and D2D rounds. • Empirical results show significant gains over baseline SD‑FL and classic gradient tracking. • Supports scalable, star‑free architectures suitable for device‑to‑device fog deployments.
Article Summaries:
- A new semi‑decentralized federated learning method, SD‑GT, has been proposed to address scalability issues in fog‑based networks that lack a conventional star topology. SD‑GT removes the need for gradient‑diversity assumptions by adding tracking terms to device updates at both the device‑to‑device (D2D) and device‑server stages. The authors derive convergence bounds for non‑convex, convex, and strongly‑convex problems and use these bounds to tune subnet sampling rates and D2D rounds, balancing performance and efficiency. Numerical tests on several datasets show that SD‑GT improves trained model quality and reduces communication cost compared with existing SD‑FL and gradient‑tracking baselines.
Sources: