Research On Decoding Method Of Implantable Brain Computer Interface Oriented To Ultrasmall Sample Set

In recent years,with the rapid development of science and technology,neural interfaces have made great progress in helping people with physical disabilities to restore their athletic ability.Among them,the implantable brain-computer interface collects electrical signals emitted by neurons by implanting neural microelectrodes into the cerebral cortex.Compared with other neural interfaces,it can collect neural signals with a higher signal-to-noise ratio and richer information.Therefore,it has received extensive attention,and many research results have been obtained in clinical trials.However,the neural signal collected by the implantable brain-computer interface will show instability over time.Therefore,the neural signal decoder often needs to be recalibrated before it can be used.This makes the implantable brain-computer interface in clinical applications.It is often accompanied by a lengthy recalibration process,which brings great inconvenience to users.In response to this problem,this article mainly studies the decoding method of the implantable brain-computer interface when facing ultra-small sample sets,aiming to achieve the decoding of neural signals by using ultra-small sample sets,thereby reducing the need for new samples for decoder recalibration.First,for the scene where there are only a few labeled samples in the target domain,a cluster based re-weighting domain adaptation algorithm is proposed.The algorithm estimates the similarity between the source domain sample and the target domain sample based on the purity of the cluster to which the source domain sample belongs and the distance between the sample and the cluster center,and assigns different weights to each source domain sample,so that the decoder can effectively use a large number of Historical data to complete the decoder recalibration with only a very small number of target domain samples.Experimental results show that the algorithm effectively reduces the need for new samples for decoder recalibration,shortens the time required for decoder recalibration,and improves the practicality of implantable brain-computer interfaces in clinical applications.Secondly,for the scene where there are very few unlabeled samples in the target domain,a marginal distribution based multi-source unsupervised domain adaptation algorithm is proposed.This algorithm assigns different weights to the base classifiers of each source domain by estimating the degree of similarity of the edge distributions between domains,and then integrates them to obtain the classifiers.Experimental results show that the algorithm completes the calibration task of the decoder with only a small number of unlabeled target domain samples,thereby eliminating the complicated steps of collecting sample labels and optimizing the process of decoder recalibration.Finally,for scenarios where no target domain samples are available,a multi-source reweighting domain transfer component analysis based domain generalization algorithm is proposed.This method uses the idea of domain generalization,based on a multi-source migration component analysis algorithm,and trains a large amount of historical data from multiple source domains to generate a decoder with better generalization ability.Experimental results show that the decoder does not need to be recalibrated before each use,and can directly predict a completely unknown target domain.This research explored the application of domain generalization algorithm in implantable brain-computer interface for the first time,and enriched the application scenarios of implanted brain-computer interface.