Digital Semantic Device-Edge Co-Inference With Task-Oriented ARQ

This paper considers a device-edge co-inference system for image classification tasks. The system splits a deep learning model between an edge device (ED) and an edge server (ES), where the ED feeds raw images to its local model and transmits the output semantic features to the ES for classification. We address two key challenges in transmitting continuous features through modern digital wireless communication systems. First, we devise a smooth sine-based surrogate function to approximate the non-differentiable quantization computation applied to the continuous feature vector during model training. This approach results in stable gradient backpropagation and efficient training convergence. Second, we propose a novel task-oriented automatic repeat request (ARQ) mechanism, namely TARQ, to overcome communication error. Compared to the conventional bit-error detection based ARQ, TARQ makes retransmission decisions based on the current SNR (Signal-to-Noise Ratio) and inference result, effectively reducing the retransmission probability without degrading the inference accuracy. Simulation results show that the proposed scheme enjoys fast training convergence and attains an inference accuracy only 0.62% lower than the ideal performance. Meanwhile, it reduces the inference delay of conventional ARQ scheme by 25.5% on average, achieving an efficient tradeoff between inference accuracy and delay.