Multi-Kernel Temporal and Spatial Convolution for EEG-Based Emotion Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) analysis for brain-computer interfaces (BCIs) is challenging due to feature extraction complexities.
• Convolutional Neural Networks (ConvNets) are used for EEG tasks but require careful architectural design for optimal feature representation.

Purpose of the Study:

• To propose a novel EEG-based emotion classification model, the Multi-kernel Temporal and Spatial Convolution Network (MultiT-S ConvNet).
• To enhance feature extraction by incorporating multi-scale kernels and separable convolutions for improved representational ability.

Main Methods:

• Developed MultiT-S ConvNet utilizing multi-scale kernels for diverse time resolutions and separable convolutions for spatial patterns.
• Integrated a lightweight gating mechanism to enhance temporal and spatial filters.
• Validated the model on DEAP and SEED EEG emotion datasets using subject-dependent and independent experiments.

Main Results:

• MultiT-S ConvNet demonstrated superior classification accuracy compared to existing methods on both DEAP and SEED datasets.
• The model achieved high accuracy with a reduced number of trainable parameters.
• The multi-scale temporal filtering module effectively extracted a broad spectrum of EEG representations, from short- to long-wavelength components.

Conclusions:

• MultiT-S ConvNet offers an effective approach for EEG-based emotion classification, outperforming current methods.
• The proposed multi-scale module can enhance the learning capacity of various EEG-based convolutional networks.
• The model's ability to capture diverse temporal EEG features holds promise for advancing BCI applications.