Machine Learning Methods Applied Research In The Brain - Machine Interface Technology

Brain-computer interfaces (BCI) systems create a novel communication channel that directly translates human intentions reflected by a brain signal into sequences of control commands for an output device such as a computer application. Therefore they could provide a new communication and control option for paralyzed patients. This will greatly enhance the ability of these subjects to manage external event and improve their living quality. BCI technology is an interdisciplinary technology integrating neuroscience, signal collection, signal processing, pattern recognition, control theory and more other relevant techniques. In this work, we mainly focused on quantitative neurophysiological investigations of human brain activaties and developed several machine learning algorithms to improve the efficiency of BCI. I have considered the following three main points:Firstly, based on support vector machines (SVM), we present a framework for single-trial online classification of imaginary left and right hand movements. The time-frequency information is extracted from two frequency bands (μandβrhythms) of EEG data with Morlet wavelets, and the SVM framework is used for accumulation of the discrimination evidence over time to infer user's motor intention. This algorithm improved the single-trial online classification accuracy as well as stability. Second, based on the neurophysiological knowledge, we have developed a discriminative spatial patterns (DSP) method for the feature extraction of movement related potentials (MRPs), and it is integrated with common spatial patterns (CSP) to extract the features from the EEG recorded during voluntary left versus right finger movement tasks. The results show that the combined spatial filters can improve the single-trial EEG classification effectively. Finally, developed was a transductive support vector machines (TSVM) algorithm to deal with the small training set problem for reducing the training process in BCI. The advantage of the transductive algorithm is the incorporation of unlabelled data during the training processes. The validity of the transductive algorithm is testified by an application to the classification of mental imagery tasks. The results show that the training efforts can be reduced significantly without increasing the classification error.