부산대학교 도서관

The large amount of data is one challenge in electroencephalogram (EEG) analysis, in which the channel Two Dimensional (2-D), time One Dimensional (1-D), and spectral (1-D) are generally considered. Convolutional neural networks (CNNs) have drawn much attention for automatic feature learning in various fields. Meanwhile, many studies have demonstrated integration from multiple sources and decisions could boost performance. However, CNN for EEG analysis usually involves millions of parameters, which easily leads to overfitting. A new model of a multistream Three Dimensional (3-D) CNN with parameter sharing is proposed for EEG. Two EEG data sets: 1) the lane-keeping task (LKT) data set and 2) sleep data set are applied. For the LKT data set, the proposed multistream 3-D CNN with parameter sharing model achieves 0.5486 root-mean-square error (RMSE), showing improvement by at least 2.77% compared to the other approaches. In the sleep data set, the error rate of the proposed model was 24.65%, showing at least 10.28% improvement in performance compared to the other methods. The lower RMSE and error rate show that the multistream 3-D CNN with parameter sharing model efficiently extracts significant features from EEG data. Moreover, the sharing mechanism even reduces the risk of overfitting and the number of parameters by comprehending common representations among multiple streams.