DSE-Mixer: A pure multilayer perceptron network for emotion recognition from EEG feature maps

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Emotion recognition from electroencephalography (EEG) signals is complex.
• Convolutional Neural Networks (CNNs) struggle with global spatial information due to local bias.
• Existing methods often limit the accuracy of EEG-based emotion recognition.

Purpose of the Study:

• To introduce the Dual-scal EEG-Mixer (DSE-Mixer) model for improved EEG feature map processing.
• To enhance the utilization of global spatial information in EEG signals.
• To achieve higher accuracy in emotion recognition tasks.

Main Methods:

• Designed the DSE-Mixer model with brain region and electrode mixer layers for multi-scale EEG information fusion.
• Implemented alternating row and column mixing for cross-regional and cross-channel communication.
• Incorporated a channel attention mechanism for adaptive channel importance learning.

Main Results:

• Achieved 95.19% arousal and 95.22% valence accuracy on the DEAP dataset.
• Reached high accuracies for four-class emotion classification on DEAP (e.g., HVHA: 92.12%).
• Obtained 93.69% average accuracy for three emotions on the SEED dataset.

Conclusions:

• DSE-Mixer demonstrates outstanding performance in emotion recognition.
• The model achieves high accuracy with significantly less computational complexity compared to CNN and Vision Transformer (ViT).
• DSE-Mixer offers a novel, compact, and effective solution for brain map processing in emotion recognition.