• Combines Mixture‑of‑Experts (MoE) with satellite‑terrestrial networks (STN) to overcome data scarcity and compute limits in federated learning. • Introduces EMS‑FL, an expert‑driven model splitting scheme that assigns only relevant experts to each edge device cluster. • Uses non‑overlapping expert assignments to enable asynchronous local training, reducing communication overhead during intermittent satellite links. • Devices upload local expert parameters to satellite only when connected, allowing efficient aggregation and model updates. • Theoretical convergence analysis demonstrates faster convergence and higher accuracy versus conventional federated learning. • Extensive experiments on public datasets and large models confirm EMS‑FL’s superiority in training speed and performance.
Article Summaries:
- Summary
A new federated learning framework, EMS‑FL, tackles the dual challenges of limited data and computational resources on edge devices while exploiting satellite‑terrestrial networks (STNs). By combining mixture‑of‑experts (MoE) models with expert‑driven model splitting, EMS‑FL assigns each device cluster only the experts most relevant to its local data, avoiding overlap. Devices train their assigned experts asynchronously and upload parameters to a satellite only during connectivity windows, reducing communication overhead. The authors provide theoretical convergence guarantees and demonstrate, through experiments on public datasets and large models, that EMS‑FL achieves faster convergence and higher accuracy than conventional federated learning approaches.
Sources: