Scalable Multi-task Semantic Communication System with Feature Importance Ranking

Jiangjing Hu, Feng‐Yu Wang, Wenjun Xu, Hui Gao, Ping ZhangICASSP 2023

Abstract: Semantic communications are expected to be an innovative solution to the emerging intelligent applications in the era of connected intelligence. In this paper, a novel scalable multitask semantic communication system with feature importance ranking (SMSC-FIR) is explored. Firstly, the multi-task correlations are investigated by a joint semantic encoder to extract relevant features. Then, a new scalable coding method is proposed based on feature importance ranking, which dynamically adjusts the coding rate and guarantees that important features for semantic tasks are transmitted with higher priority. Simulation results show that SMSC-FIR achieves performance gain w.r.t. individual intelligent tasks, especially in the low SNR regime.

Introduction

Traditional communication systems focus on data-level fidelity, aiming to transmit every bit perfectly. However, as intelligent applications proliferate, the focus is shifting toward semantic communication, which prioritizes the transmission of the "meaning" required for specific downstream tasks. This paradigm shift can significantly reduce data volume and bandwidth requirements. Despite this potential, existing semantic systems often suffer from rigidity—being optimized for single tasks—and a lack of scalability. When signal-to-noise ratios (SNR) drop, these systems often fail to degrade gracefully, either losing critical information or requiring computationally expensive parallel architectures for multiple tasks.

System Architecture

To address these challenges, we propose the Scalable Multi-task Semantic Communication system with Feature Importance Ranking (SMSC-FIR). The architecture leverages multi-task correlations to improve both efficiency and robustness.

Architecture of the proposed SMSC-FIR system.

The system utilizes a Joint Semantic Encoder that extracts features shared across multiple tasks simultaneously, such as vehicle re-identification (ReID) and attribute classification. This exploits the fact that correlated tasks often rely on similar visual features. A key innovation is the Scalable Feature Selector, which employs Feature Importance Ranking (FIR). By using Gradient-weighted Class Activation Mapping (Grad-CAM), the system identifies which features are most critical for task performance. Based on real-time Channel State Information (CSI) and SNR, the system dynamically selects the top-ranked features to fit within the available transmission budget.

Experimental Evaluation

The SMSC-FIR system was evaluated using the VeRi-776 dataset, a large-scale benchmark for vehicle recognition. The researchers focused on three representative tasks: vehicle ReID, vehicle color classification, and vehicle type classification. The system was compared against traditional methods (JPEG + LDPC) and other deep learning-based scalable baselines using random or sequential feature selection.

Performance comparison of SMSC-FIR across ReID, color, and type classification tasks.

As shown in Figure 6, SMSC-FIR demonstrates superior performance, particularly in low SNR regimes. At 0dB SNR, the system achieved a 40.0% improvement in ReID accuracy over sequential selection and a 212.2% improvement over random selection. For classification tasks (color and type), the system's accuracy approached the "complete transmission" upper bound even at low SNRs, proving that a small subset of high-importance features is often sufficient for high-accuracy classification. While ReID is a more complex task requiring more features to reach peak performance, SMSC-FIR consistently remained the top performer across all tested SNRs.

Conclusion

The SMSC-FIR framework effectively exploits multi-task correlations and semantic prioritization to provide a scalable communication solution. By ranking features by their semantic contribution, the system ensures robust performance under adverse channel conditions without the need for retraining. Future work will focus on reducing the computational overhead of gradient calculations for FIR and exploring dynamic task weighting to further enhance system flexibility in real-world traffic scenarios.

Summary updated

Abstract

Semantic communications are expected to be an innovative solution to the emerging intelligent applications in the era of connected intelligence. In this paper, a novel scalable multitask semantic communication system with feature importance ranking (SMSC-FIR) is explored. Firstly, the multi-task correlations are investigated by a joint semantic encoder to extract relevant features. Then, a new scalable coding method is proposed based on feature importance ranking, which dynamically adjusts the coding rate and guarantees that important features for semantic tasks are transmitted with higher priority. Simulation results show that SMSC-FIR achieves performance gain w.r.t. individual intelligent tasks, especially in the low SNR regime.
BibTeX
@inproceedings{hu2023scalable,
  author = {Jiangjing Hu and Feng‐Yu Wang and Wenjun Xu and Hui Gao and Ping Zhang},
  title = {{Scalable Multi-task Semantic Communication System with Feature Importance Ranking}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
}
Figure