SigCLR: A contrastive learning approach to unsupervised modulation recognition and novelty detection

We introduce a contrastive learning architecture for radio datasets called SigCLR. We show how this method of unsupervised machine learning can be applied for a variety of tasks including but not limited to modulation recognition and novelty detection. We show results for a functioning modulation class detector, where an unknown signal can be passed into the trained network and it can be compared to a number of known modulations.