Deep Representation-Based Domain Adaptation for Nonstationary EEG Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) data exhibit significant inter-subject variability, hindering classifier performance across different individuals.
• Collecting extensive subject-specific data for motor imagery tasks is time-consuming and impractical.

Purpose of the Study:

• To develop an end-to-end deep domain adaptation method for enhancing single-subject EEG classification performance.
• To leverage knowledge from multiple source subjects to improve classification accuracy on a target subject.

Main Methods:

• A novel deep domain adaptation method jointly optimizing a feature extractor, classifier, and domain discriminator.
• Utilizing adversarial learning for domain adaptation and center loss to reduce intra-subject non-stationarity.
• Mapping raw EEG signals into a discriminative latent feature space for improved classification.

Main Results:

• The proposed method demonstrated efficacy in improving classification performance on target subjects by utilizing data from source subjects.
• Experiments on BCI Competition IV datasets (IIa and IIb) validated the method's effectiveness.
• The approach successfully reduced feature distribution shifts between source and target domains.

Conclusions:

• The developed deep domain adaptation method shows promise for reducing BCI calibration time.
• This technique facilitates more efficient and accurate BCI system development and deployment.
• The method enables accurate label prediction in the target domain by leveraging consistent deep features from both domains.