Self-Supervised EEG Denoising via Intelligent Partitioning
A new method called Intelligent Partitioning for Self-supervised Denoising (iPSD) enables training of deep learning EEG denoisers without requiring clean reference signals. Classical signal processing fails on wearable EEG due to time-varying artifacts, while deep learning typically needs unobtainable artifact-free data. iPSD partitions an input EEG segment into independent noisy realizations sharing the same underlying signal, allowing self-supervision even in zero-shot settings with a single segment. The method is validated on wearable EEG denoising tasks.
Key facts
- iPSD eliminates need for clean reference EEG
- Learns to partition input EEG into independent noisy realizations
- Enables self-supervised deep learning denoising
- Works in zero-shot settings with single EEG segment
- Addresses time-varying pervasive artifacts in wearable EEG
- Classical methods fail due to fixed or heuristic rules
- Deep learning shows promise but requires unobtainable artifact-free data
- Validated on wearable EEG denoising
Entities
Institutions
- arXiv