Spike Sorting With Clustering Ensemble Based On Local Weighting

Neurons are the basic unit of the brain.Monitoring the activity of individual neuron is helpful to understand the working mechanism of the brain.The collected neuron recordings are the mixture of multiple neurons.They can be separated by spike sorting which consists of spike detection,feature extraction and clustering.This dissertation mainly studies the clustering of spike sorting.Based on weighted voting and spectral clustering algorithm,the idea of local weighting is introduced to improve the performance of the two algorithms in spike sorting.The details are as follows:Firstly,the framework of spike sorting is introduced,which includes spike detection,feature extraction and clustering.The electronic signal emitted by every neuron has its own unique spike.Due to this characteristic,spike sorting can be used to separate the spikes and further to achieve the separation of neurons.Spike detection can extract the spikes of neuron recordings.Then,feature extraction can be used to extract the features of spikes.Clustering algorithm can be used to cluster the feature space and finally the spike sorting is completed.Secondly,a locally weighted voting algorithm is proposed.The weighted voting method is a clustering ensemble algorithm,which has the characteristics of simple implementation and high efficiency.However,the weights in the traditional weighted voting method is distributed uniformly to the clustering members.The unreliability of every cluster in the clustering members is not be taken into account.So the performance of clustering will be reduced.In this dissertation,a locally weighted voting is proposed,in which Shannon entropy is introduced to evaluate the unreliability of clusters and the clusters are weighted according to their unreliability.Experimental results show that the proposed locally weighted voting is superior to traditional weighted voting.Thirdly,a locally weighted co-association matrix-based spectral clustering is proposed.Spectral clustering can work on any datasets and obtain globally optimal solution.However,Gaussian kernel is used to construct the similarity matrix in spectral clustering.It is difficult to adjust the parameter and select the suitable distance measurement.Based on the idea of clustering ensemble,locally weighted co-association(LWCA)matrix can fuse the information of each partition to improve the accuracy and stability of clustering.Meanwhile,it can avoid computing the distance between the sample points.In this dissertation,spectral clustering based on LWCA matrix is proposed,which adopts LWCA matrix to construct the similarity matrix.Experimental results show that spectral clustering based on LWCA matrix is superior to the traditional spectral clustering.Finally,locally weighted voting and locally weighted co-association matrix-based spectral clustering algorithms are applied in spike sorting,respectively.Experimental results show that the accuracy of spike sorting based on the locally weighted co-association matrix-based spectral clustering algorithm is superior to spike sorting based on the locally weighted voting algorithm.However,the former consumes more time than the latter.