Reproduce “Self-Contrastive Learning based Semi-Supervised Radio Modulation Classification” on OpenResearch

About this paper

Dongxin Liu, Peng Wang, Tianshi Wang, Tarek Abdelzaher

Researchers at the University of Illinois Urbana-Champaign and the US Army Research Laboratory introduced SemiAMC, a semi-supervised framework that employs self-supervised contrastive learning for radio modulation classification. This method effectively leverages unlabeled data to achieve high classification accuracy, demonstrating significant performance gains with limited labeled samples and sometimes exceeding fully supervised baselines.

Key insights

• Self-supervised contrastive learning can effectively extract robust and discriminative features from unlabeled radio signals, significantly reducing the reliance on labeled data for AMC.
• The proposed encoder architecture, combining CNN layers for spatial features and LSTM layers for temporal dependencies, is well-suited for learning representations from sequential I/Q signal data.
• Even when ample labeled data is available, the features learned through contrastive pre-training can lead to intrinsically more robust and superior classification performance compared to training a supervised model from scratch.

Summary by alphaXiv.