Task-Oriented Semantic Communications With Efficient Retransmission Using Feature Selection

Abstract

Task-oriented semantic communications transmits task-relevant features, instead of raw data, to save bandwidth resources and reduce communication latency, which is of great attention in the Internet-of-Things (IoT), smart healthcare, and autonomous driving. In task-oriented semantic communications, efficient retransmission is essential for reducing semantic errors and preventing the waste of communication resources. In this thesis, we propose an efficient retransmission scheme for task-oriented semantic communications, leveraging a deep neural network (DNN) to optimize the processing of task-relevant information. We introduce a novel feature selection method to transmit more relevant and less redundant feature elements selectively. In addition, we design a DNNbased ACK/NAK generator, which detects semantic errors based on the classification with no cyclic redundancy check (CRC), thereby reducing communication overhead. Simulation results show that the proposed model outperforms a baseline model by improving communication efficiency and classification accuracy, which demonstrates the effectiveness of selective transmission of features by our proposed method.