A Self-Supervised Method for Accelerated Training of Neural Communication Receivers

Self-supervised learning (SSL), which is a branch of unsupervised learning, is a new machine learning paradigm for learning from large unlabeled datasets. In this paper we apply principles of SSL to the channel autoencoder problem from communications theory. We demonstrated this by first performing an SSL pre-training step using a contrastive loss, the training time of a neural receiver can be significantly reduced, even when the extra pre-training time has been accounted for. This approach could be used to improve the performance of neural receivers in a wide variety of channel conditions.