Wireless Data Acquisition for Edge Learning: Importance-Aware

Abstract

By deploying machine learning algorithms at the network edge, edge learning recently emerges as a promising framework to support intelligent mobile services. It effectively leverages the rich data collected by abundant mobile devices, and exploits the proximate edge computing resource for low-latency execution. Edge learning crosses two disciplines, machine learning and wireless communication, and thereby gives rise to many new research issues. In this paper, we address a wireless data acquisition problem, which involves a retransmission decision in each communication round to optimize the data quality-vs-quantity tradeoff. A new retransmission protocol called importance-aware automatic-repeat-request (importance ARQ) is proposed. Unlike classic ARQ focusing merely on reliability, importance ARQ selectively retransmits a data sample based on its uncertainty that can be measured using the model under training. Underpinning the proposed protocol is an elegant communication-learning relation between two corresponding metrics, i.e., signal-to-noise ratio (SNR) and data uncertainty. This new relation facilitates the design of a simple threshold based policy for retransmission decisions. As demonstrated via experiments with real datasets, the proposed method avoids learning performance degradation caused by channel noise while achieving faster convergence than conventional SNR-based ARQ.