Self-Supervised Learning for Brain Waves

Cut-and-paste augmentations like CutMix and CutPaste have reshaped image learning; this project investigates how their spirit can be reimagined for multichannel time series such as EEG. The aim is self-supervised pretraining that teaches models to recover spatio-temporal–spectral context: windows are cut, shifted, or recombined across channels and trials to create puzzles whose solutions require understanding spatial relationships, inter-channel coherence, and rhythm structure.

The study identifies which SSL objectives—masked modeling, contrastive prediction, patch forecasting—derive the greatest benefit from temporal CutPaste-style operations, and why. Particular care is taken to preserve physiological plausibility while still forcing hard context tasks. Success is defined by improved downstream performance on subject-independent classification, anomaly detection, and event detection, together with interpretable attention to cortical regions that signal true context recovery rather than shortcut artifacts.

References:

Yun, Sangdoo, et al. "Cutmix: Regularization strategy to train strong classifiers with localizable features." Proceedings of the IEEE/CVF international conference on computer vision. 2019.

Li, Chun-Liang, et al. "Cutpaste: Self-supervised learning for anomaly detection and localization." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.

Al-Saegh, Ali, Shefa A. Dawwd, and Jassim M. Abdul-Jabbar. "CutCat: An augmentation method for EEG classification." Neural Networks 141 (2021): 433-443.

Nochumsohn, Liran, and Omri Azencot. "Data Augmentation Policy Search for Long-Term Forecasting." Transactions on Machine Learning Research.