The SCEEGNet: An Efficient Learning Method for Emotion Recognition Based on the Few Channels

Emotion recognition based on Electroencephalogram (EEG) is an essential task in the advanced stage of artificial intelligence. Two major limitations of the prior approaches are: (1) the features of brain laterality are not fully exploited, which results in inefficient learning; (2) too many channels are used to make the model hard to migrate and deploy. In the current study, we proposed a symmetric EEG channel convolutional neural network (SCEEGNet) using 8 EEG channels to solve the problems mentioned above. Specifically, we selected 4 pairs of symmetric EEG channels and reconstructed them into a Symmetric Channel Topology (SCT) which is a cross arrangement of left and right brain regions. After that, the SCEEGNet captured the features of the brain laterality via convolving dual EEG channels. The SCEEGNet was lightweight and had an end-to-end pipeline. The validity of the SCEEGNet was tested using the SEED dataset. The results showed that the SCEEGNet reached an average accuracy of 98.37% for classifying three emotions using only 8 EEG channels' data, demonstrating the efficacy of our approach.