Dynamic Joint Domain Adaptation Network for Motor Imagery Classification.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalogram (EEG) is crucial for brain-computer interfaces (BCI) but requires lengthy calibration.
• Current EEG classification methods rely on handcrafted features or extensive session-specific annotated data.
• This limits the practical application and widespread adoption of EEG-based BCIs.

Purpose of the Study:

• To develop a novel dynamic joint domain adaptation network for improved EEG classification.
• To address the limitations of time-consuming calibration and the need for annotated samples in EEG-based BCI.
• To learn domain-invariant feature representations for enhanced classification performance in target domains.

Main Methods:

• A dynamic joint domain adaptation network utilizing adversarial learning.
• Employing a global discriminator for marginal distribution alignment across domains.
• Utilizing a local discriminator to reduce conditional distribution discrepancy and a dynamic adversarial factor for adaptive alignment importance.

Main Results:

• The proposed method demonstrated superior performance in EEG classification compared to state-of-the-art techniques.
• Achieved effective domain-invariant feature representation learning.
• Successfully leveraged information from source sessions to improve target domain classification.

Conclusions:

• The novel dynamic joint domain adaptation network effectively enhances EEG classification performance.
• The method significantly reduces the reliance on time-consuming calibration procedures.
• This approach holds promise for more efficient and practical EEG-based BCI applications.