Adding self-supervised learning to neural decoders lets you use unlabeled brain recordings to improve performance, which is crucial when labeled behavioral data is limited—a common real-world constraint in neurotechnology applications.
This paper introduces MOJO, a training framework that combines self-supervised learning (masked autoencoding) with supervised learning to improve neural decoders for brain-computer interfaces. By leveraging unlabeled neural data alongside labeled behavioral data, MOJO achieves better performance than traditional supervised-only approaches, especially when labeled data is scarce.