A Self-Supervised Task-Agnostic Embedding for EEG Signals

Brain-Computer Interfaces (BCIs) have great potential for improving the lives of people with disabilities. The success of a BCI system is largely driven by the accuracy of the BCI decoder. This accuracy, in turn, may be limited by the amount of labelled training data available for supervised machine learning algorithms. The success of deep learning algorithms in other computer science areas has not reached the field of BCI decoding due to this lack of abundant labelled data. We use a novel deep learning architecture trained in a self-supervised manner to learn a common vector representation (embedding) of EEG signals that can be used in different BCI tasks. The vector representation is trained using EEG recordings without using any task labels. We validate our embedder using two separate BCI tasks: seizure detection and motor imagery, and assess its usefulness through distance similarity metrics in a clustering approach. The derived embeddings were successful in distinguishing binary classes in both tasks.