.png){kind=icon}
.png){kind=icon}
Return
On the Effectiveness of Task-Oriented Digital Semantic Communication System
.png){kind=icon}
PRE
AI
DOI：10.1109/TCCN.2025.3628475.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Task-oriented digital semantic communication systems utilize neural networks to extract and transmit task-relevant features, which significantly improves the communication efficiency compared to traditional communication systems. These systems also ensure compatibility with modern digital communication frameworks, and thus attract much attention in emerging resource-constrained communication scenarios. However, existing research has limitations in optimizing the information representation efficiency of features extracted at the transmitter, thereby constraining the system’s effectiveness as measured by inference performance such as classification accuracy. In this paper, two regularization terms are proposed to address this limitation. First, a transmission effectiveness regularization term is introduced to enhance the downstream task performance by reducing the redundant information in the extracted features. Second, a semi-orthogonal regularization term is proposed to optimize the vector quantization module, which improves quantization efficiency and model generalization by maximizing codeword separation and minimizing linear correlations between codewords. These proposed regularization terms are then incorporated into the objective function of an existing robust digital semantic communication system. Furthermore, a new implementation scheme and a corresponding optimization objective is presented. Experimental results demonstrate that, the presented scheme achieves better inference performance with low latency and exhibits stronger robustness in the case of mismatched training and testing channels compared to the baseline schemes.
Keywords:
Task-oriented communication
digital semantic communication
joint source-channel coding
feature learning
Journal
I
IF:
7
Papers:
1.5K
Citations:
5.5K