MI-DABAN: A dual-attention-based adversarial network for motor imagery classification

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalography (EEG)-based brain-computer interfaces (BCIs) offer convenience and safety.
• Distributional disparities in EEG signals hinder direct use of cross-subject data for classifier training.
• Existing domain transfer and adversarial learning methods align data globally but neglect class boundaries and domain-specific information.

Purpose of the Study:

• To propose a novel dual-attention-based adversarial network (MI-DABAN) for motor imagery classification.
• To enhance single-subject BCI performance by leveraging multi-subject knowledge.
• To address limitations of existing methods, such as feature blurring and loss of domain-specific information.

Main Methods:

• Developed a novel adversarial learning approach using iterative maximization and minimization of classifier output differences for domain alignment without extra discriminators.
• Implemented two unshared attention blocks to preserve domain-specific shallow features, preventing negative transfer.
• Validated the MI-DABAN framework on BCI Competition IV Datasets 2a and 2b.

Main Results:

• The proposed MI-DABAN effectively improves single-subject motor imagery classification performance.
• The dual-attention mechanism successfully preserves crucial domain-specific information.
• Experimental results demonstrate superior performance compared to existing methods on public EEG datasets.

Conclusions:

• The MI-DABAN framework offers an effective solution for cross-subject EEG-based BCI classification.
• The novel adversarial learning and attention mechanisms enhance domain alignment and preserve essential features.
• This approach shows significant potential for advancing motor imagery classification in BCI applications.