eduzhai > Applied Sciences > Engineering >

Subject-Aware Contrastive Learning for Biosignals

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 17 pages

Abstract: Datasets for biosignals, such as electroencephalogram (EEG) andelectrocardiogram (ECG), often have noisy labels and have limited number ofsubjects (<100). To handle these challenges, we propose a self-supervisedapproach based on contrastive learning to model biosignals with a reducedreliance on labeled data and with fewer subjects. In this regime of limitedlabels and subjects, intersubject variability negatively impacts modelperformance. Thus, we introduce subject-aware learning through (1) asubject-specific contrastive loss, and (2) an adversarial training to promotesubject-invariance during the self-supervised learning. We also develop anumber of time-series data augmentation techniques to be used with thecontrastive loss for biosignals. Our method is evaluated on publicly availabledatasets of two different biosignals with different tasks: EEG decoding and ECGanomaly detection. The embeddings learned using self-supervision yieldcompetitive classification results compared to entirely supervised methods. Weshow that subject-invariance improves representation quality for these tasks,and observe that subject-specific loss increases performance when fine-tuningwith supervised labels.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×