Modeling And Identification Of Brain Motor Neural System

Brain-computer interface (BCI) is a new technology to realize the brain communication and controlling of the external environment. With the development of the multi-channel neural signal recoding and computer control technology, the decoding algorithms which extract movement information from the brain cortical neural ensemble activity is a key component to relate the brain signal with external environment. The aim of this dissertation is to identify and model the brain motor neural system. For this purpose, the decoding algorithms that extract movement information from the motor cortical neural spike trains and the methods to analyze the neural signals from the view of temporal coding are researched.Firstly, the methods to model the relation between motor cortical neural signals and arm movement directions are researched. A multi-class support vector machines (SVM) algorithm of a binary tree recognition strategy is used to predict the movement direction with the firing rate patterns of neural ensemble. The performance of the SVM based neural activity recognition is compared with that of the linear population vector algorithm (PVA) and the learning vector quantization (LVQ) approach. The results show that the SVM method has better learning and generalization performance, which demonstrates that the SVM algorithm is a suitable approach for neural signals analyses. In addition, the least squares support vector machines (LS-SVM) method is also used to model the brain motor neural system. The LS-SVM algorithm not only has good performance similar to SVM, but also costs less computational time than SVM. That demonstrates the LS-SVM method is more suitable for online analyses of neural signals, and it holds hope for a possibly more accurate BCI for neural prosthesis.Secondly, a nonlinear ARX (NARX) model based on LS-SVM is established to identify the relationship between the firing rate patterns of cortical neural ensemble and 3D hand positions. The results show that nonlinear NARX methods prevail against linear ARX method to model the motor cortical neural system. And the LS-SVM algorithm has higher prediction accuracy and better generalization performance than the ANN approach to build the nonlinear model. The best combinations of neurons are also selected from the entire neural ensemble for modeling. The results show that the model built with less neuron can achieve better performance. That can lead to BCI system demanding lower computational power.To analyze the temporal patterns in neural spike trains, the spiking neural networks (SNN) which propagate information by the timing of individual spikes are studied in this dissertation. The spiking neural model, network architecture, simulation issues, and learning algorithm of SNN are systematically introduced. Two methods are presented to improve the SpikeProp algorithm, an error back-propagation (BP) learning rule suited for SNN. One method is to use the adaptive learning rates with momentum to speed up the convergence and dynamic performance of the SNN. Another is to present a more biologically plausible spiking response model (SRM) to describe the spiking neurons by not neglecting the dependence of the postsynaptic potential upon the firing times of the postsynaptic neuron, and to derive an additional BP learning rule for the coefficient of the refractoriness function. These improvements make the SNN in which neurons can spike multiple times can transfer information more efficiently.Then the SNN method is proposed to extract movement direction and target orientation from the temporal pattern of the motor cortical neural spike trains directly. A one-layer and a two-layer feedforward SNN are used to analyze the activity of motor cortical neurons. The results show that the SNN algorithm is a feasible to analyze the timing of spike trains, and the multiple-layer network has higher computational power. On the other hand, the comparison results of the temporal pattern recognition by the SNN algorithm and the firing rate analysis by the ANN approach are consistent with the recent development about the neural coding that temporal coding is more biologically plausible than the rate coding. The SNN method is promising to extract more useful movement information from the neural spike trains without temporal information lost.Finally, the summary of the dissertation and the future work to be investigated are presented.