Task-Oriented Low-Label Semantic Communication With Self-Supervised Learning

TL;DR

This work addresses efficient task-oriented semantic communication under limited labeled data by proposing SLSCom, a two-stage framework that first pretrains a semantic encoder on abundant unlabeled data using self-supervised, contrastive signals aligned with an information bottleneck objective, and then jointly trains the JSCC and semantic decoder with a small labeled set. The approach leverages a classification pretext (InfoNCE) to maximize task-relevant information and a reconstruction pretext to minimize residual, enabling robust end-to-end inference over multipath OFDM channels with few labels. Extensive simulations on CIFAR10, SVHN, and Flowers show SLSCom consistently outperforms conventional digital coding, training-from-scratch, and pre-trained-transfer baselines, particularly in low-SNR and low-label scenarios, and remains resilient to data distribution shifts and partial label missing. The results highlight the practical impact of combining self-supervised semantic extraction, IB-based feature selection, and joint task-oriented transmission for edge-enabled visual analytics in wireless networks, with potential extensions to multi-device settings.

Abstract

Task-oriented semantic communication enhances transmission efficiency by conveying semantic information rather than exact messages. Deep learning (DL)-based semantic communication can effectively cultivate the essential semantic knowledge for semantic extraction, transmission, and interpretation by leveraging massive labeled samples for downstream task training. In this paper, we propose a self-supervised learning-based semantic communication framework (SLSCom) to enhance task inference performance, particularly in scenarios with limited access to labeled samples. Specifically, we develop a task-relevant semantic encoder using unlabeled samples, which can be collected by devices in real-world edge networks. To facilitate task-relevant semantic extraction, we introduce self-supervision for learning contrastive features and formulate the information bottleneck (IB) problem to balance the tradeoff between the informativeness of the extracted features and task inference performance. Given the computational challenges of the IB problem, we devise a practical and effective solution by employing self-supervised classification and reconstruction pretext tasks. We further propose efficient joint training methods to enhance end-to-end inference accuracy over wireless channels, even with few labeled samples. We evaluate the proposed framework on image classification tasks over multipath wireless channels. Extensive simulation results demonstrate that SLSCom significantly outperforms conventional digital coding methods and existing DL-based approaches across varying labeled data set sizes and SNR conditions, even when the unlabeled samples are irrelevant to the downstream tasks.

Figures (12)

• Figure 1: Diagram of the proposed semantic communication.
• Figure 2: The framework and network overview of SLSCom. The parameters of fully-connected (FC) layers are denoted in the format of (input size, output size). The parameters of (transposed) convolutional layers are represented in the format of filters/stride/padding.
• Figure 3: A Venn diagram for (P2) with self-supervision.
• Figure 4: Pre-training for semantic encoder with pretext tasks.
• Figure 5: Comparison of test accuracy on CIFAR10 versus the number of labeled samples.