Mental workload recognition from EEG signals via semi-supervised autoencoders.

Area of Science:

• Cognitive Science
• Neuroscience
• Machine Learning

Background:

• Mental workload assessment is crucial in human-computer interaction, system design, and healthcare.
• Electroencephalogram (EEG)-based workload recognition commonly uses supervised learning, which struggles with limited labeled data.

Purpose of the Study:

• To develop a novel semi-supervised autoencoder framework for robust EEG-based mental workload recognition.
• To leverage abundant unlabeled EEG data to overcome the limitations of scarce labeled datasets.

Main Methods:

• Proposed a semi-supervised autoencoder integrating supervised objectives with unsupervised autoencoders.
• Developed a joint function to minimize both data reconstruction and prediction errors, enhancing model discriminative power.
• Incorporated skip connections between autoencoder layers to mitigate vanishing/exploding gradient issues.

Main Results:

• The proposed framework achieved high accuracy in binary mental workload classification.
• Demonstrated effective utilization of both labeled and unlabeled EEG data.
• The model showed improved performance compared to traditional supervised methods on two distinct EEG datasets.

Conclusions:

• Semi-supervised autoencoders offer a promising approach for EEG-based mental workload recognition.
• The integration of supervised and unsupervised learning with architectural enhancements effectively addresses data scarcity challenges.
• The framework provides a scalable and accurate solution for real-world cognitive effort monitoring.