EEGNet-based multi-source domain filter for BCI transfer learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning models for brain-computer interfaces (BCI) struggle with inter-individual differences in electroencephalography (EEG) data.
• Training deep learning models for BCI typically requires substantial data per subject, increasing costs.

Purpose of the Study:

• To propose a novel transfer learning method, EEGNet-MDFTL, to reduce training data needs and improve BCI performance.
• To address the challenge of inter-individual variability in EEG data for deep learning applications.

Main Methods:

• Developed EEGNet-MDFTL, a transfer learning approach utilizing bagging ensemble learning.
• Employed a multi-source domain filter that leverages model loss value to select relevant data sources.
• Focused on learning domain-invariant features from multiple data sources.

Main Results:

• Achieved a decoding accuracy of 91.96%, outperforming baseline and state-of-the-art methods.
• Demonstrated sustained high accuracy even with data reduced to 1/8th of the original amount.
• Confirmed that the source domain filter effectively selects similar domains to boost model accuracy.

Conclusions:

• EEGNet-MDFTL significantly improves EEG decoding performance using limited data, thereby reducing BCI training costs.
• The proposed method highlights the effectiveness of ensemble learning in extracting domain-invariant features for robust BCI models.