Self-supervised pretraining

Self-supervised pretraining learns useful representations from large amounts of unlabeled neural data by solving a pretext task whose targets come from the data itself — predicting a masked segment, contrasting temporally near versus far windows, or forecasting future activity — then fine-tuning the learned encoder on a small labeled downstream task. It directly attacks BCI's defining bottleneck: labels (calibrated, task-locked trials) are scarce and expensive, whereas raw EEG or iEEG is abundant.

Representative approaches include relative-positioning and temporal-shuffling pretext tasks for EEG, the contrastive wav2vec-style BENDR, and masked spike reconstruction for intracortical data. The consistent finding is improved data efficiency and cross-subject transfer, but gains are dataset-dependent and can evaporate when the pretraining corpus differs from the target in montage, hardware or task — pretraining is not a substitute for handling distribution shift.