Progressive Transmission of High-Dimensional Data Features for Inference at the Network Edge

Abstract

Uploading high-dimensional features from edge devices to an edge server over wireless channels creates a communication bottleneck for edge inference. To tackle the challenge, we propose the progressive feature transmission (ProgressFTX) protocol, which minimizes the overhead by progressively transmitting features until a target confidence level is reached. The control of the protocol to accelerate inference is designed with two key operations. The first, importance-aware feature selection, guides the server to select the most discriminative feature dimensions. The second is transmission-termination control such that the feature transmission is stopped when the incremental uncertainty reduction by further transmission is outweighed by its communication cost. The indices of the selected features and transmission decision are fed back to the device in each slot. The sub-optimal policy is obtained for classification using a convolutional neural network. Experimental results on a real-world dataset shows that ProgressFTX can substantially reduce the communication latency compared to conventional feature pruning and random feature transmission.