Related Experiment Video
Updated: Jun 21, 2025
Author Spotlight: Enhancing Neurorehabilitation Through EEG, Motor Imagery, and Virtual Reality
Published on: May 10, 2024
Unsupervised and semi-supervised domain adaptation networks considering both global knowledge and prototype-based
Dongxue Zhang1, Huiying Li1, Jingmeng Xie2
1Jilin University, College of Computer Science and Technology, Changchun, Jilin Province, China; Key Laboratory of Symbol Computation and Knowledge Engineering, Jilin University, Changchun 130012, China.
This study introduces GPL, a domain adaptation method for electroencephalography (EEG) signals, improving motor imagery classification by aligning global and local data features. The method enhances brain-computer interface accuracy in unsupervised and semi-supervised settings.
Area of Science:
- Neuroscience
- Machine Learning
- Biomedical Engineering
Background:
- Non-stationarity in electroencephalography (EEG) signals causes session variability, hindering model development and data sharing.
- Accurate classification of motor imagery signals is crucial for advancing brain-computer interfaces (BCIs).
Purpose of the Study:
- To propose a novel domain adaptation method, GPL, for enhancing motor imagery signal classification accuracy.
- To address the challenges of EEG signal variability and limited labeled target data.
Main Methods:
- GPL employs a dual approach: global feature space alignment using Maximum Mean Difference (MMD) loss and prototype-based local class information via memory banks and contrastive losses.
- Unsupervised and semi-supervised versions are implemented, with an entropy-aware strategy in unsupervised settings to manage pseudo-label confidence.
- Source contrastive loss organizes source features, while interactive contrastive loss promotes cross-domain information exchange.
Main Results:
- Experiments on BCI Competition IV Dataset IIa and GigaDB yielded average classification accuracies of 86.03% and 84.22%, respectively.
- The proposed method demonstrates significant improvements in EEG decoding for motor imagery tasks.
Conclusions:
- GPL is an effective EEG decoding model that enhances motor imagery classification by addressing signal non-stationarity through domain adaptation.
- The findings contribute to the development of more robust and reliable motor imagery-based brain-computer interfaces.
