Deep Source Semi-Supervised Transfer Learning (DS3TL) for Cross-Subject EEG Classification.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) translate electroencephalogram (EEG) signals into device commands.
• Training reliable EEG recognition models typically requires extensive labeled data, which is time-consuming and user-unfriendly.
• Semi-supervised learning (SSL) and transfer learning offer strategies to reduce the dependency on labeled data by utilizing unlabeled or auxiliary data.

Purpose of the Study:

• To propose Deep Source Semi-Supervised Transfer Learning (DS3TL) for EEG-based BCIs.
• To reduce the amount of labeled data required for training reliable EEG recognition models for new subjects.
• To leverage unlabeled data from a source subject and unlabeled data from a target subject for improved classifier training.

Main Methods:

• DS3TL integrates a hybrid SSL module (pseudo-labeling and consistency regularization), a weakly-supervised contrastive module (using true and pseudo-labels), and a domain adaptation module (uncertainty reduction).
• The source subject has a small amount of labeled and a large amount of unlabeled EEG trials.
• All EEG trials from the target subject are unlabeled.

Main Results:

• DS3TL outperformed a supervised learning baseline that used significantly more labeled training data.
• DS3TL demonstrated superior performance compared to state-of-the-art SSL approaches when using the same amount of labeled data.
• Experiments were conducted on three diverse EEG datasets from different tasks.

Conclusions:

• DS3TL is the first approach in EEG-based BCIs to effectively utilize unlabeled source data for enhanced target classifier training.
• The proposed method significantly reduces the burden of data labeling for individual users.
• DS3TL offers a promising direction for developing more efficient and user-friendly BCIs.