Data augmentation for EEG

Data augmentation synthesizes additional training examples by applying label-preserving transformations, a first-line remedy for the small trial counts of BCI. EEG-appropriate transforms include additive noise, amplitude scaling, time shifting and time warping, channel dropout, band-stop filtering (frequency masking), Fourier-surrogate phase randomization, mixup between trials, and physiologically motivated schemes such as small rotations of the sensor montage. Cropped training — sampling many overlapping sub-windows per trial — is a widely used implicit augmentation.

The essential caveat is that EEG lacks the near-universal invariances of natural images: a left-right flip is meaningful in vision but changes the class in a lateralized motor-imagery task, and aggressive amplitude or frequency perturbation can erase the ERD or SSVEP feature being decoded. Effective augmentation must respect the paradigm's known invariances, and its benefit tends to shrink as labeled data or self-supervised pretraining grows.