Applying Self-Supervised Representation Learning for Emotion Recognition Using Physiological Signals.

Area of Science:

• Affective Computing
• Machine Learning
• Artificial Intelligence

Background:

• Current emotion recognition models rely on supervised learning, requiring large, expertly labeled datasets.
• Data annotation is costly and impractical, especially in healthcare settings demanding specialized knowledge.
• This limits the scalability and accessibility of emotion recognition technologies.

Purpose of the Study:

• To explore the application of self-supervised learning (SSL) for developing emotion recognition methods.
• To investigate SSL's potential to learn representations from unlabeled physiological signals for affective state classification.
• To compare the performance of SSL against traditional supervised learning in emotion recognition tasks.

Main Methods:

• The study introduces key concepts of emotions and SSL methodologies for affective state recognition.
• A convolutional neural network was trained using both self-supervised and fully supervised approaches.
• Experimental analysis was conducted using three distinct emotion datasets.

Main Results:

• Self-supervised representations demonstrated the ability to learn broadly applicable features from unlabeled data.
• SSL significantly improved data efficiency and showed strong transferability across datasets.
• The performance of SSL was competitive with fully supervised methods, without needing labeled data for initial learning.

Conclusions:

• Self-supervised learning is a viable and efficient paradigm for emotion recognition in affective computing.
• SSL reduces the dependency on large labeled datasets, making emotion recognition more practical in data-scarce domains like healthcare.
• SSL-trained models offer competitive performance and enhanced data efficiency compared to supervised approaches.