Contrastive Learning Consistent and Identifiable Latent Embeddings for EEG

Extracting informative EEG data into low-dimension latent embeddings is important for storing and analyzing these neuron signals and applying them to various applications, such as modern human-computer interaction (HCI) techniques. We use the contrastive learning algorithm on time-domain features of EEG in both discovery-driven (self-supervised) and hypothesis (supervised) manners to encode the EEG data into latent embeddings that are proven consistent and identifiable. The self-supervised embeddings have the potential to be used for a range of downstream tasks, while the supervised embeddings have very high decoding accuracy for specific tasks. With embeddings encoded from EEG features collected within every 0.5-s window, the accuracy of recognizing the identities of persons by decoding the self-supervised and supervised embeddings is as high as 96.2% and 99.6%, respectively. Our method and results can promote new HCI techniques, e.g., automatically connecting users to their roles in AR games once they wear EEG-capable devices. The source code is available at: https://www.github.com/liangfengsid/timeEegContrastive .