Self-Supervised EEG Emotion Recognition Models Based on CNN.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Emotion plays a vital role in human life.
• Electroencephalogram (EEG) signal analysis for emotion classification is gaining traction due to advancements in brain-computer interface (BCI) and machine learning.
• Current fully-supervised methods for EEG-based emotion classification are resource-intensive.

Purpose of the Study:

• To investigate the application of self-supervised learning methods for improving resource efficiency in EEG-based emotion classification.
• To develop and evaluate a deep multi-task convolutional neural network (CNN) using a self-supervised approach for emotion recognition.

Main Methods:

• A self-supervised approach was employed to train a deep multi-task CNN.
• Six signal transformations were applied to unlabeled EEG data to create a pretext task for training.
• The trained network's convolutional layers were frozen, and the fully connected layer was adapted for emotion recognition using labeled EEG data.
• Experiments were conducted on the SEED and DEAP affective datasets to assess performance across different data scales.

Main Results:

• Self-supervised learning effectively captures internal data representations.
• The proposed method demonstrated significant savings in computation time compared to fully-supervised learning.
• The self-supervised approach showed improved performance in EEG-based emotion classification.

Conclusions:

• Self-supervised learning offers a more resource-efficient alternative for EEG-based emotion classification.
• This methodology can enhance the performance of emotion classification models compared to conventional fully supervised techniques.
• The findings suggest a promising direction for developing efficient and effective BCI systems.