Emotion Recognition Based on Electroencephalogram Using Semi-supervised Generative Adversarial Network.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) based emotion recognition is gaining traction due to its non-invasive nature.
• A significant challenge is the scarcity of labeled EEG data, limiting classifier accuracy.
• Semi-supervised learning offers a potential solution by leveraging abundant unlabeled data.

Purpose of the Study:

• To propose and evaluate a Semi-supervised Generative Adversarial Network (SGAN) for improving EEG-based emotion recognition accuracy.
• To demonstrate the effectiveness of SGAN in utilizing large unlabeled EEG spectrogram datasets.
• To address the limitations imposed by small labeled datasets in emotion recognition.

Main Methods:

• Implementation of a Semi-supervised Generative Adversarial Network (SGAN) architecture.
• Training the discriminator network to distinguish between true labeled, true unlabeled, and synthetic EEG spectrogram data.
• Utilizing an 80%:20% validation split on EEG spectrogram data from 50 participants for binary emotion classification (positive/negative valence).

Main Results:

• The proposed SGAN method achieved an accuracy of 84.83% with only 50% labeled data.
• This accuracy surpasses the baseline discriminator network (83.5%) and prior studies (around 78%).
• Ablation study confirmed the SGAN architecture's superiority over a mere discriminator network.

Conclusions:

• SGAN effectively enhances EEG emotion recognition by enabling the discriminator to learn from both labeled and unlabeled data.
• The ability to distinguish true from synthetic samples forces feature focus, improving generalization and accuracy.
• SGAN presents a viable solution for overcoming data scarcity in EEG-based emotion recognition research.