Data Importance-Assisted Multi-User Scheduling in MIMO Edge Learning Systems

With the wide development of intelligent communication systems, efficient data transmission is critical to fast edge learning in multi-user multiple-input multiple-output (MIMO) systems since the data acquisition from massive edge devices has become a bottleneck. To cope with the mismatch between the empirical probability of the transmitted data and the expected one, this paper first proposes to quantify data importance using the Kullback-Leibler divergence. Then, we design a multi-user scheduling criterion that combines the channel state information and data importance indicators, followed by an iterative multi-user scheduling algorithm. Finally, experimental results demon-strate that the proposed multi-user scheduling strategy signifi-cantly improves the learning efficiency and the test accuracy of edge learning systems.