Multi-Source Transfer Learning for EEG Classification Based on Domain Adversarial Neural Network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) classification is crucial but challenged by inter-subject variability.
• Collecting sufficient labeled EEG data for individual subjects is time-consuming and labor-intensive.
• Existing models often struggle with performance degradation due to these individual differences.

Purpose of the Study:

• To propose a novel multi-source transfer learning method for robust EEG classification.
• To address the challenge of inter-subject variability in EEG data.
• To improve the performance and generalizability of EEG classifiers.

Main Methods:

• A domain adversarial neural network (DANN) was designed, incorporating a feature extractor, classifier, and domain discriminator.
• A unified multi-source optimization framework was developed to leverage data from multiple sources.
• The DANN aims to minimize domain shift between different subjects or datasets.

Main Results:

• The proposed method demonstrated significant advantages in EEG classification across three public datasets.
• The multi-source optimization framework improved classification performance through weighted predictions.
• The domain adversarial approach effectively reduced the impact of individual differences.

Conclusions:

• The developed multi-source transfer learning method offers a promising solution for EEG classification.
• Domain adversarial neural networks are effective in mitigating domain shift in EEG data.
• The approach enhances classifier performance by integrating knowledge from multiple source domains.