Stochastic weight averaging enhanced temporal convolution network for EEG-based emotion recognition

Abstract

Emotion plays an important role in human–computer interactions, in which emotion recognition is the key problem. Electroencephalogram (EEG) has been widely used in emotion recognition due to its high temporal resolution and reliability. In this study, we propose a strategy in which stochastic weight averaging is introduced into an improved temporal convolutional network for emotion recognition. The temporal convolution network not only is suitable for sequence models such as the recurrent neural networks, but also retains the characteristics of parallel computing similar to convolutional neural networks. Considering that the traditional softmax loss does not explicitly encourage discriminative learning of features, we further introduce an improved version of the loss function that explicitly encourages intraclass compactness and interclass separability between learned features. Moreover, the method of stochastic weight averaging is introduced into the network framework to make the network find the point closest to the global optimum by adjusting the learning rate and updating the weight, which can effectively alleviate the local optimum problem. We test the performance of the proposed strategy on two open emotion EEG datasets: DEAP and SEED. The experimental results show that the new strategy has higher recognition accuracy than the state-of-the-art approaches. Moreover, to investigate the general pattern of brain functional connectivity in different individuals, we select the key electrodes and analyze the roles of various brain regions in processing different emotions.