Feature Importance-Aware Task-Oriented Semantic Transmission and Optimization
Yining Wang, Shujun Han, Xiaodong Xu, Haotai Liang, Rui Meng, Chen Dong, Ping Zhang
Abstract
Incorporating both semantic level information and effectiveness level performance, the task-oriented semantic communication system has been designed for various tasks of different datatype. Although semantic communication improves the spectral utilization to some extent, indiscriminate transmission of semantic information for task-oriented semantic communication can still result in waste of wireless resources. In this paper, we propose an importance-aware joint source-channel coding (I-JSCC) framework for task-oriented semantic communications. A joint semantic-channel transmission (JSCT) mechanism is designed by selectively transmitting task-important features to reduce communication overhead. We define a new metric named task-oriented semantic spectral efficiency (TOSSE) to evaluate the effectiveness and efficiency of the proposed system, which measures the effective semantic information carried by each semantic symbol. An importance-aware semantic resource allocation problem is formulated to maximize the total TOSSE of all users by jointly optimizing the channel assignment and feature selection vector. To solve this problem, a knowledge-assisted proximal policy optimization (K-PPO) based reinforcement learning (RL) algorithm is proposed. The experimental results conducted on CIFAR100 dataset demonstrate the efficacy of the K-PPO algorithm, while also highlighting the superiority of the importance-aware semantic communication system in terms of the TOSSE.