An Investigation of Deep Learning Models for EEG-Based Emotion Recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Emotion recognition from electroencephalography (EEG) signals is crucial for understanding human responses.
• Traditional machine learning methods for EEG-based emotion recognition often require complex, expert-driven feature extraction.
• Deep learning offers an end-to-end approach, utilizing raw signal features and time-frequency spectrums.

Purpose of the Study:

• To investigate the effectiveness of various deep learning models for EEG-based emotion recognition.
• To compare the performance of Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM model.
• To analyze model performance based on accuracy, training speed, and parameter sensitivity.

Main Methods:

• Utilized the DEAP dataset for experiments.
• Implemented and evaluated DNN, CNN, LSTM, and CNN-LSTM models.
• Conducted comparative analysis of model performance, including accuracy and training parameters (epochs, learning rate, dropout).

Main Results:

• CNN and CNN-LSTM models achieved high classification accuracy, reaching 94.17% and 90.12% respectively on raw EEG data.
• DNN models demonstrated faster training speeds but lower accuracy compared to CNN and CNN-LSTM.
• LSTM models showed less stability and slower convergence, especially with comparable parameter counts.

Conclusions:

• CNN and CNN-LSTM models are highly effective for EEG-based emotion recognition, offering superior accuracy and feature extraction.
• Deep learning models significantly advance the field by automating feature extraction from complex EEG data.
• Model selection involves trade-offs between accuracy, training efficiency, and stability, with CNN-LSTM showing promising results.