By deploying machine learning algorithms at the network edge, edge learningrecently emerges as a promising framework to provide intelligent services. Itcan effectively leverage the rich data collected by abundant mobile devices andexploit the proximate edge computing resource for low-latency execution. Edgelearning crosses two disciplines, machine learning and wireless communication,and thereby gives rise to many new research issues. In this paper, we address awireless data acquisition problem, which involves a retransmission decision ineach communication round to optimize the data quality-vs-quantity tradeoff. Anew retransmission protocol called importance-aware automatic-repeat-request(importance ARQ) is proposed. Unlike classic ARQ focusing merely onreliability, importance ARQ selectively retransmits a data sample based on itsuncertainty which helps learning and can be measured using the model undertraining. Underpinning the proposed protocol is an elegantcommunication-learning relation between two corresponding metrics, i.e.,signal-to-noise ratio (SNR) and data uncertainty. This new measure facilitatesthe design of a simple threshold based policy for retransmission decisions. Asdemonstrated via experiments with real datasets, the proposed method avoidslearning performance degradation caused by channel noise while achieving fasterconvergence than conventional SNR-based ARQ.