Hidden-layer visible deep stacking network optimized by PSO for motor imagery EEG recognition

A novel method called PSO optimized hidden-layer visible deep stacking network (PHVDSN) is proposed for feature extraction and recognition of motor imagery electroencephalogram (EEG) signals. A prior knowledge is introduced into the intermediate layer of deep stacking network (DSN) and the hidden nodes are expanded by the unsupervised training of restricted Boltzmann machine (RBM) for the parameter initialization. Then particle swarm optimization (PSO) is applied to optimize the input weights, aiming at alleviating the risk of being immersed in the curse of dimensionality. The performance of the proposed method is evaluated with real EEG signals from different subjects. Experimental results show that the recognition accuracy of PHVDSN is superior to some state-of-the-art feature extraction algorithms. Furthermore, on another benchmark data set where the EEG sessions for each subject are recorded on separated days, the proposed method is demonstrated to be robust against transferring from session to session.