Self-Supervised Learning Enabled Task-Oriented Semantic Communication Using Limited Labels

Deep learning (DL)-based semantic communication generally relies on extensive labeled data to cultivate the essential semantic knowledge for semantic extraction and interpretation. In this paper, we propose a self-supervised learning-based task- oriented semantic communication (SLSCom) framework, tailored for scenarios with limited access to labeled data. Specifically, we establish a task-relevant semantic encoder by employing self- supervision and the information bottleneck (IB) principle on unlabeled data, aiming to strike a balance between differential entropy and inference performance. Given the computational challenges in the IB-based problem, we introduce two pretext tasks based on the self-supervision, providing a tractable lower bound. SLSCom facilitates the interpretation of the task- relevant information using limited labeled data, ensuring the robust inference performance on downstream tasks. We validate the effectiveness of the proposed SLSCom through simulation experiments conducted on image classification tasks.