Latency-Driven Joint Feature Extraction and Resource Allocation for Multi-Task Multi-Access Semantic Communications

IEEE Journal on Selected Areas in Communications, vol. 43, pp. 3984–3999

Abstract

Semantic communication has achieved great progress in improving efficiency for completing tasks successfully, instead of directly transmitting bits. However, substantial challenges remain in real-time intelligent communication, which demands stringent low latency and rapid understanding of massive data. In this paper, we propose a latency-driven design for promoting real-time multi-task multi-access semantic communications. More specifically, we investigate a deep learning-based framework for multi-access scenarios, where multiple users with individual latency requirements continuously request real-time semantic updates from an edge server. Two typical image-based semantic tasks, i.e., image classification and object detection, are considered as representative multi-task example. Furthermore, since the low-latency requirements in real-time systems force the application of finite blocklength (FBL) codes to be a significant consideration, we take into account the effects of FBL on transmission reliability. To adapt to the low-latency demands, we adopt a parameter-sharing strategy for multi-task computer vision (CV) applications and design an adaptive mixed-precision compression module for effective feature compression. The design target is to maximize the minimum weighted task success probability among all users via jointly optimizing feature extraction, mixed-precision quantization bit selection, transmit power allocation and semantic decoding. To facilitate the overall joint optimization, we propose an approach for efficient optimal decision-making on joint quantization bit selection and power allocation, which is integrated into deep learning process for adaptive feature extraction. Simulation results verify the promising performance of our proposed latency-driven design for real-time multi-task CV applications, as well as the superior benefits of our proposed efficient optimal resource allocation for real-time communication scheduling.

Authors 3

1. Yulin Hu

   Wuhan University

   Affiliation as printed

   School of Electronic Information, Wuhan University, Wuhan, China

2. Xiaopeng Yuan Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   Chair of Information Theory and Data Analytics, RWTH Aachen University, Aachen, Germany

3. Anke Schmeink Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   Chair of Information Theory and Data Analytics, RWTH Aachen University, Aachen, Germany