Decentralized Online Learning With Compressed Communication for Near-Sensor Data Analytics

Abstract

Near-sensor data analytics advocates processing data locally near their sources, rather than gathering them for centralized processing. It can reduce communication costs and is particularly suitable for networked sensor systems whose data are geo-distributed. As most sensors and associated devices have limited computing power, it is desirable for them to collaborate, especially in a decentralized manner, so that the workload can be distributed and no single point becomes a bottleneck. In this letter, we present a decentralized machine learning algorithm with communication compression capability that can serve as the core of a near-sensor data analytics task. Owing to its online nature and reduced communication overhead, the proposed method is particularly suitable for real-world sensor network systems with energy and bandwidth constraints.