Task-Oriented Feature Re-Encoding and Forwarding for Relay-Assisted Device-Edge Co-Inference

In this paper, we explore a task-oriented relay strategy for enhancing classification task performance in device-edge co-inference systems. In particular, we propose a deep neural network (DNN) based Re-encoding and Forwarding (RF) strategy, whereby the relay re-encodes the extracted feature from the edge device (ED) and then delivers it to the edge server (ES) for task execution. Unlike the conventional amplify and forward (AF) strategy where the extracted feature is simply amplified and relayed in fixed dimension, RF can produce re-encoded features with flexible dimensions to maximize the inference accuracy under constrained communication resources. Simulation results show that, unlike AF, the optimal relay position of RF is no longer at the midpoint of link ED-ES. For all considered channel conditions, RF can provide up to 11.6% higher task inference accuracy over AF by jointly adjusting the relay position and the feature dimension at the relay device (RD).