Research On Neural Decoding Based On Machine Learning For Intracortical Brain-machine Interfaces

With the continuous development of computer science,biology,control theory and other disciplines,brain-machine interface(BMI)has shown great potential in the clinical application of motor function reconstruction for disabled people.BMI can be divided into non-intracortical brain-machine interfaces and intracortical brain-machine interfaces(i BMI)by the different ways of acquiring neural recordings.This paper mainly focuses on the intracortical brain-machine interface.i BMI acquires neural signals by intracortical neural electrodes,and then interprets the neural signals of the brain into the control commands of external devices.This transformation realizes the information interaction between the brain and external devices and decodes the motor intentions,which provides a new way for the disabled to reconstruct motor function and has an important prospect in future clinical application.This technology involves many interdisciplinary subjects and has become one of the spotlights for scientific research.Firstly,this paper conducted an i BMI experiment with non-human primates.We designed a monkey grasp experimental paradigm,and then identified the precise location to implant electrodes in the monkey brain and implanted three neural electrodes surgically,obtaining the neural signals in the brain which are highly related to grasping movement.Finally,the obtained original neural signals were preprocessed,and acquired the spike trains by the threshold crossing method,and the neural activity matrix(NAM)composed of the feature vectors corresponding to the activity of a single neuron was eventually obtained.To solve the problem that many neural signals in BMI are difficult to label,this paper proposed a greedy sampling active learning algorithm.Different from the traditional method which randomly selects the training data,this method actively selects the most diverse samples in the training set to label and then trains the decoder.To solve the problem that greedy sampling is easily affected by the outliers in the raw data,we proposed to denoise the neural signals and improve the robustness and accuracy of the decoder by detecting outliers,so as to achieve higher stability and accuracy than the popular active learning algorithms.Finally,the performance of the proposed algorithms is verified by the neural signals extracted from the macaque grasp experiment,and the average decoding accuracy of the proposed three algorithms is 94%,97% and 98% respectively,realizing the neural decoding of grasping motion direction of macaque.It is necessary for BMI to decode online in clinical applications.However,on the one hand,the nonstationarity of neural recordings leads to decoder's recalibration while applicating.On the other hand,disabled people are unable to provide real motor output to train a decoder.As a result,it prohibits a decoder from meeting the requirements of online decoding in the clinical.To solve this problem,a reinforcement learning algorithm combining transfer learning and the mini-batch method was proposed in this paper.Compared with conventional decoding methods,this method uses the reward signal from the interaction with the environment rather than body motor output and only uses a few new current samples to update weights,improving the domain adaptation between the historical data and current data in the meantime.Eventually,the accuracy of the decoder is improved to more than 90%,and the efficiency of weights updating is improved by more than 70 times.