Parallel convolutional-linear neural network for motor imagery classification

Deep learning, recently, has been successfully applied to image classification, object recognition and speech recognition. However, the benefits of deep learning and accompanying architectures have been largely unknown for BCI applications. In motor imagery-based BCI, an energy-based feature, typically after spatial filtering, is commonly used for classification. Although this feature corresponds to the estimate of event-related synchronization/desynchronization in the brain, it neglects energy dynamics which may contain valuable discriminative information. Because traditional classiication methods, such as SVM, cannot handle this dynamical property, we proposed an architecture that inputs a dynamic energy representation of EEG data and utilizes convolutional neural networks for classification. By combining this network with a static energy network, we saw a significant increase in performance. We evaluated the proposed method and compared with SVM on a multi-class motor imagery dataset (BCI competition dataset IV-2a). Our method outperforms SVM with static energy features significantly (p < 0.01).