Few-Shot Adaptive Learning for Robust Task-Oriented Semantic Communication

Wenjie Han, Anbang Zhang, Chenyuan Feng, Ruxiao Chen, Shuang Zhang, Shuaishuai Guo

Keywords: Task-oriented semantic communications, few shot learning, channel adaption

TL;DR: Using few-shot learning for channel adaption in task-oriented semantic communication

Abstract: Task-oriented semantic communication (ToSC) emerges as a promising approach for executing remote inference tasks. While existing ToSC systems are generally trained under specified channel conditions, the volatile nature of real-world channel conditions poses significant adaptation challenges to ToSC. To this end, we propose an adaptive ToSC system for dynamic environments via few-shot learning in this paper. The method utilizes the data-driven mechanism named vector quantized variational autoencoder (VQ-VAE) to dynamically optimize the codebook and generate non-uniform modulation codebooks that are closely aligned with specific task objectives. In addition, few-shot learning and transfer learning techniques are adopted to facilitate efficient learning on small datasets, allowing the system to swiftly adjust its operating parameters to adapt to new communication conditions. Experimental results show that the proposed method achieves superior performance compared to traditional channel-adaptive methods, especially in environments with low signal-to-noise ratios (SNR).

Submission Number: 10