Learning-efficient Transmission Scheduling for Distributed Knowledge-aware Edge Learning.

Edge learning is a promising enabler to leverage the distributed local data for powering the artificial intelligence at the edge network. Moreover, incorporating the external domain knowledge into purely data-driven learning models can further enhance the performance. In this paper, by taking both the benefits of edge learning and knowledge fusion, we propose a novel distributed knowledge-aware edge learning framework, in which the edge devices individually train the learning models with the assistance of the local knowledge bases at the edge devices and the global knowledge base at the edge server. Due to the limited cache capability, the edge device can only cache a small-scale local knowledge base, which restricts the performance gain by local knowledge fusion. Meanwhile, uploading local data from multiple edge devices for global knowledge fusion may lead to the air-interface congestion. To overcome these issues, we first formulate the global loss decay maximization problem with transmission scheduling decisions. Specifically, we derive the closed-form relationship between transmission scheduling and the learning performance. Then, we depict the implicit relationship between the knowledge fusion and the global loss decay via establishing a specific multi-armed bandit (MAB) framework, and derive an asymptotically-optimal solution accordingly. Extensive simulations demonstrate that the proposed policies outperform the state-of-art policies.