Learning Generalized Representations of EEG between Multiple Cognitive Attention Tasks.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Cognitive tasks such as Stroop, Eriksen Flanker, and Psychomotor Vigilance Task (PVT) measure attention.
• These tasks, despite content differences, rely on visual attention.
• Electroencephalogram (EEG) data offers insights into brain activity during attention tasks.

Purpose of the Study:

• To investigate the ability of machine learning models to learn generalized representations from EEG data across different cognitive attention tasks.
• To compare the performance of Support Vector Machine (SVM), EEGNet, and TSception in classifying attention from EEG data.
• To determine if commonalities exist in EEG patterns across diverse attention tasks.

Main Methods:

• Conducted intra-task and inter-task attention classification experiments using EEG data from Stroop, Flanker, and PVT tasks.
• Employed Support Vector Machine (SVM), EEGNet, and TSception as classification models.
• Utilized cross-validation techniques for intra-task performance evaluation.

Main Results:

• TSception achieved superior classification accuracies in intra-task experiments: 82.48% (Stroop), 88.22% (Flanker), and 87.31% (PVT).
• Deep learning methods (EEGNet, TSception) demonstrated better performance than SVM in inter-task experiments.
• Accuracy drops in inter-task deep learning classification were generally not significant, suggesting robust generalized representations.

Conclusions:

• Common neural representations for cognitive attention tasks can be learned from EEG data.
• Deep learning models are effective in capturing these generalized representations across different attention tasks.
• Findings support the potential for using EEG and deep learning for understanding and potentially diagnosing attention-related conditions.