Importance-Aware Data Pre-Processing and Device Scheduling for Multi-Channel Edge Learning

The large-scale deployment of intelligent Internet of things (IoT) devices have brought increasing needs for computation support in wireless access networks. Applying machine learning (ML) algorithms at the network edge, i.e., edge learning, requires efficient training, in order to adapt themselves to the varying environment. However, the transmission of the training data collected by devices requires huge wireless resources. To address this issue, we exploit the fact that data samples have different importance for training, and use an influence function to represent the importance. Based on the importance metric, we propose a data pre-processing scheme combining data filtering that reduces the size of dataset and data compression that removes redundant information. As a result, the number of data samples as well as the size of every data sample to be transmitted can be substantially reduced while keeping the training accuracy. Furthermore, we propose device scheduling policies, including rate-based and Monte-Carlo-based policies, for multi-device multi-channel systems, maximizing the summation of data importance of scheduled devices. Experiments show that the proposed device scheduling policies bring more than 2% improvement in training accuracy.