Identification of emotional valences via memory-informed deep neural network with entropy features

Emotion plays an important role in people's everyday routine and work. Using electroencephalograph (EEG) signals to identify emotional states of human brain is one of the most valuable methods for emotion recognition. This paper studied positive and negative emotional valences identification from EEG signals via memory-informed deep neural network with entropy features. To quantify EEG signals over time, we first used sliding time windows to calculate sample entropy in EEG signals. Then we integrated a life-long memory module into deep neural network to accumulate prior knowledge of the entropy features of positive and negative emotional valences during training phase, so as to enhance the performance of emotional valences identification. Finally, we performed our experimental analysis with the SEED (SJTU Emotion EEG Dataset) dataset, a publicly available EEG dataset for emotion analysis. The average accuracy of 92.22% was achieved for the identification of positive and negative emotional valences for 15 subjects in SEED dataset. The experimental results showed that the proposed framework could effectively achieve the identification of positive and negative emotional valences from EEG signals, which had broad application prospects in healthcare decision-making system.