CV-EEGNet: A Compact Complex-Valued Convolutional Network for End-to-End EEG-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Electroencephalogram (EEG)-based emotion recognition commonly uses real-valued neural networks, primarily focusing on the time-amplitude domain.
• These real-valued networks may not fully capture crucial frequency, phase, and amplitude characteristics inherent in EEG signals, potentially limiting emotion decoding accuracy.
• Amplitude-phase coupling and spectral structures are vital for effective emotion recognition from EEG data.

Purpose of the Study:

• To introduce complex-valued neural networks for the first time in EEG-based emotion recognition.
• To propose a novel end-to-end architecture, Complex-valued EEGNet (CV-EEGNet), designed to leverage complex-valued representations.
• To investigate the potential of CV-EEGNet in capturing intricate signal characteristics for improved emotion decoding.

Main Methods:

• Raw EEG signals are transformed into complex-valued spectra using the Fast Fourier Transform.
• The CV-EEGNet architecture employs complex-valued spectral, spatial, and depthwise-separable convolution modules.
• These modules extract frequency structures, spatial topologies, and semantic representations while maintaining amplitude-phase relationships, followed by a complex-valued classifier.

Main Results:

• Experiments on the SEED (three-class) and SEED-IV (four-class) datasets demonstrated the effectiveness of the proposed CV-EEGNet method.
• t-SNE visualizations confirmed the enhanced discriminability of the representations learned by the complex-valued network.
• The results indicate superior performance compared to existing real-valued approaches.

Conclusions:

• Complex-valued neural networks show significant potential for raw-signal EEG-based emotion recognition.
• The proposed CV-EEGNet architecture effectively utilizes amplitude-phase information for more accurate emotion decoding.
• This work opens new avenues for leveraging complex-valued models in affective computing and brain-computer interfaces.