A Dual-Adversarial Model for Cross-Time and Cross-Subject Cognitive Workload Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) signals are crucial for cognitive workload decoding (CWD).
• Existing CWD methods struggle with subject-independent and long-term scenarios, showing decreased accuracy.
• Current solutions often require large datasets or lack robust feature extraction and clear category distinctions.

Purpose of the Study:

• To propose a novel Bi-Classifier Joint Domain Adaptation (BCJDA) model for robust EEG-based CWD.
• To address challenges in cross-time and cross-subject CWD accuracy deterioration.
• To develop a model that overcomes limitations of existing CWD approaches.

Main Methods:

• Developed a BCJDA model comprising a feature extractor, domain discriminator, and Bi-Classifier.
• Employed adversarial processes for domain-wise and class-wise feature alignment.
• Utilized cross-gradient difference maximization within the Bi-Classifier for enhanced performance.

Main Results:

• The BCJDA model demonstrated superior performance in EEG-based CWD tasks.
• Achieved significant improvements in recognizing low, medium, and high cognitive workload levels.
• The proposed model effectively handles cross-time and cross-subject variations.

Conclusions:

• The BCJDA model offers a robust solution for subject-independent and long-term EEG-based CWD.
• Adversarial domain and class alignment are effective strategies for improving CWD accuracy.
• The BCJDA model provides a promising direction for advancing CWD applications.