Research On A Spike Sorting Method And Its CUDA Implementation

Spike sorting is a fundamental preprocessing step for neuroscience studies,in which the detected spikes generated by an individual unit fall into same cluster.It’s also a crucial part in the research of the invasive brain-computer interfaces.With the development of electrodes,lots of electrodes are used simultaneously,resulting in huge workload for manual classification.An unsupervised automatic classification algorithm can solve this problem.Otherwise,the increase in number of electrodes leads to increase in time of spike sorting.Most of the feature extraction techniques used for spike sorting cann’t give the most discriminative projection subspace,leading to clusters overlapping and a poor sorting accuracy when the noise level is high.In this work,we propose a noise-robust and automatic unsupervised spike sorting algorithm named LD_DP to improve performance on classification.The algorithm combines LDA and clustering by density peak(DP)to selsct the most discriminative subspace and groups similar clusters to automatically detect the number of clusters.On condition that the algorithm is suitable for parallel,we optimize the LDA_DP algorithm with CUDA technology to achieve acceleration on GPU.By paralleling the data from each channel and designing parallel program for operations in algorithm,the time consumption is reduced.The results show that the LDA_DP algorithm can automatically classify spikes in an unsupervised way.Comparative results on publicly simulated datasets demonstrate that our algorithm is noise-robust and behaves better than PC A or discrete wavelets transform by improving cluster distinction.The average accuracy is 0.95,and the minimum accuracy is 0.86.The parallel classifier based on CUDA parallelizes the data,completes the sorting in DP with bitonic sorting network,and parallelizes all operations.Compared with the CPU-based classifier,the parallel GPU-based classifier greatly reduces computation time,eapecially when there are lots of channels.In multi-channel classification,the LDA_DP algorithm based on CUDA can get a high accuracy,reduce the burden on the experimenter and the time consumpation.