Reservoir Splitting method for EEG-based Emotion Recognition
Anubhav Anubhav, Kantaro Fujiwara · 2023
This paper presents a novel reservoir splitting method to train an efficient Reservoir Computing model for Emotion Recognition using Electroencephalogram (EEG) signals. Since different brain lobes have distinct functions and combining these functions results in the final decision, we propose splitting a single reservoir into multiple reservoirs dedicated to distinct lobes and integrating them to imitate the human brain functioning. We utilise the EEG signals from the publicly available GAMEEMO dataset for experiments. EEG signals are input to the single reservoir and proposed multiple reservoir configurations to obtain representations. Various classifiers: Ridge regression, Support Vector Machine (SVM), Gradient Boosted Classifier (GBC), and Random Forest are trained on the representations obtained from the reservoirs. We follow Leave-One-Subject-Out (LOSO) strategy to train these classifiers. Comparing the classification accuracy, we notice that representations from the proposed models of multiple reservoirs outperform the single reservoir model, and SVM performs better than the other classifiers. Furthermore, on increasing the reservoir size, the testing accuracy for all the classifiers attains a peak for both the Valence and Arousal domain. The proposed reservoir splitting method can be extended to explore diverse splitting configurations for future work.