Edge Intelligence Over the Air: Two Faces of Interference in Federated Learning
This work addresses the challenge of limited spectral resources in next-generation wireless networks for enabling scalable federated learning, though it is incremental as it provides an overview rather than a new solution.
The paper examines the dual roles of interference in federated edge learning systems that use over-the-air computations, highlighting both its positive and negative impacts on model training, scalability, and other aspects like privacy and latency.
Federated edge learning is envisioned as the bedrock of enabling intelligence in next-generation wireless networks, but the limited spectral resources often constrain its scalability. In light of this challenge, a line of recent research suggested integrating analog over-the-air computations into federated edge learning systems, to exploit the superposition property of electromagnetic waves for fast aggregation of intermediate parameters and achieve (almost) unlimited scalability. Over-the-air computations also benefit the system in other aspects, such as low hardware cost, reduced access latency, and enhanced privacy protection. Despite these advantages, the interference introduced by wireless communications also influences various aspects of the model training process, while its importance is not well recognized yet. This article provides a comprehensive overview of the positive and negative effects of interference on over-the-air computation-based edge learning systems. The potential open issues and research trends are also discussed.