Scale-Space Processing and Clustering for Efficient Multi-Electrode Data Analysis of Large-size Neuronal Ensembles

Abstract

Identifying clusters of neurons with correlated activity in large-size neuronal ensembles from high-density multielectrode array recordings is an emerging problem in computational neuroscience. A new engineering approach is proposed that relies on representing multiple neural spike trains in a scale-space in which a spectral clustering algorithm is able to identify clusters of correlated firing within different behavioral contexts (temporal bin-width choices). The method constitutes a natural extension to the multiscale representation of the neural data obtained from the array-based multiresolution spike detection and sorting algorithms previously developed. The advantage of the proposed method is its ability to efficiently identify populations of recorded neurons with correlated activity independent of the temporal scale from which rate functions are typically estimated. Moreover, it relies on simultaneously maximizing cluster aggregation based on similarity as well as cluster segregation based on dissimilarity across any number of neuronal spike trains. We compare the performance to the classical A-means and the probabilistic (Bayesian) clustering algorithms on a complex synthesized data set to illustrate the substantial gain in clustering accuracy