EEG-Based Emotion Estimate Using Shallow Fully Convolutional Neural Network with Boost Training Strategy

Emotion recognition using Electroencephalogram (EEG) has drawn the attention of many scholars. However, there are few studies looking into regressive approach. Actually, human affective states are continuous rather than discrete. This paper focuses on how to estimate continuous affective status from EEG recordings. A Shallow Fully Convolutional Network (SFCN) with Boost Training Strategy is proposed to estimate affective status, including Valence, Arousal, Dominance, and Liking. SFCN is presented to extract the emotional relative features automatically from preprocessed EEG instead of using handcrafted features. With Global Average Pooling (GAP) layer, SFCN can solve the effect of unreliability of label introduced by segmented-augmentation method. Moreover, Boost Training Strategy is designed to train model with low memory cost and further improves the performance of SFCN. Experiments on DEAP dataset demonstrate the effectiveness of proposed approaches. Results show that Mean Square Error (MSE) for Valence, Arousal, Dominance, Liking are 3.9181, 3.6009, 3.4441 and 4.806, respectively.