Deep Neural Classifiers For Eeg-Based Emotion Recognition In Immersive Environments

Abstract

Emotion recognition has become a major endeavor in artificial general intelligence applications in recent years. Although significant progress has been made in emotion recognition for music, image and video stimuli, it remains largely unexplored for immersive virtual stimuli. Our main objective for this line of investigation is to enable consistently reliable emotion recognition for virtual reality stimuli using only cheap, commercial-off-the-shelf electroencephalography (EEG) headsets which have significantly less recording channels and far lower signal resolution commonly called “Wearable EEG” as opposed to medical-grade EEG headsets with the ultimate goal of applying EEG-based emotion prediction to procedurally-generated affective content such as immersive computer games and virtual learning environments through machine learning. Our prior preliminary study has found that the use of a 4-channel, 256-Hz was indeed able to perform the required emotion recognition tasks from VR stimuli albeit at classification rates of between 65-89% classification accuracy only using Support Vector Machines (SVMs) and K-Nearest Neighbor (KNN) classifiers. For this particular study, we attempt to improve the classification rates to above 95% by conducting a comprehensive investigation into the use of various deep neural-based learning architectures for this domain. By tuning the deep neural classifiers in terms of the number of hidden layers, number of hidden nodes and the nodal dropout ratio, the emotion prediction accuracy was able to be improved to over 96%. This shows the continued promise of the application of wearable EEG for emotion prediction as a cost-effective and userfriendly approach for consistent and reliable prediction deployment in virtual reality-related content and environments through deep learning approaches.