Three class emotions recognition based on deep learning using staked autoencoder

Abstract

Emotion recognition is a hot spot in advanced humancomputer interaction system, which is of great significance in artificial intelligence, health care, distance education, military field and so on. The paper builds a stacked autoencoder deep learning classification network consist of an input layer, two autoencoder hidden layers and a softmax classifier output layer based on SJTU Emotion EEG Dataset (SEED). Pretrain the first autoencoder employed L-BFGS to optimize the cost function. Then pretrain the second autoencoder with the output of first autoencoder. Finally send to the softmax classifier. Pretrain each autoencoder in forward propagation, then fine-tuning the whole network in back propagation. The well-trained network is used to classify three emotion states including happy, neural and grief. The raw inputs are differential entropy of EEG signal in five rhythmic frequencies band and the differential entropy of whole EEG signal. Fourteen experiments are performed with 5-fold cross validation, the average classification accuracy of three class emotion states is 59.6%, 66.27%, 71.97%, 78.48%, 82.56% and 85.5%. The result shows the higher frequency band differential entropy like Gamma band is more relative to emotion reaction.