Reduction And Visualization Of Neural Information In Invasive Brain-Machine Interface

Brain-machine Interface(BMI) is one of the hottest international frontier technology involving computer, biomedicine, neuroscience and materials science. It builds a artificial communication channel between the brain and devices without going through outside nerve and muscle system. According to the way of signal acquisition, it can be divide into non-invasive BMI and invasive BMI. Compared with non-invasive BMI, the signals of invasive BMI are more informative and have more advantages over real-time and accuracy of information decoding. But the neural information from invasive BMI is more complex and of high dimension. High-dimensional neural information will affect accuracy and feasibility of decoding. The information is also hard to be understood, presented, processed and analyzed and this brings great challenges to effect resolution of brain signal.In this paper, we reduce high dimensional data by feature selection and feature extraction and research on reduction and visualization of neural information from monkey to get more effect and underlying information. The essence of reduction is feature selection. It selects features from high dimensional neural information, getting rid of redundant and noise feature. Then these selected information is used for decoding to achieve real-time control by brain. The essence of visualization is feature extraction. It applies feature extraction method to project high dimensional neural information into three dimensional space and we get the trajectory of neural activity. By the trajectory, we can unearth the characteristics of cluster and distribution hiding in high dimensional data and analyze the dynamic property of neural information.The major work of this paper is as follow. The first one is the study on neural reduction. We propose a method based on information theory and considering both redundancy and correlation. Without the decrease of generalization of model, we can use only a little neural information to decode well enough. By the reduction, the decoding speed increase much. Second we focus on the visualization of neural information. We apply a feature extraction method called laplacian eigenmap which is based on manifold learning to project high dimensional neural information into three dimensional space for visualization. By the neural trajectory from visualization we analyze the relation between neural information and body movement and excavate the dynamic characteristics of neural activity. At last, we design and implement a real-time brain control system with function of neural reduction. Neural reduction ensures the system work in real time. The system integrates function of signal acquisition, reduction, decoding, paradigm, output and feedback. It can be used for training of experiment and also for real-time brain control.