Modelos baseados em redes neurais não supervisionadas para aplicação em problemas de spike sorting

Abstract

Extracellular recordings contain neural spike signals that come from different biological neurons. The proper detection and separation of the spikes according to the neurons they originate from is usually referred to as spike sorting. The spike sorting task is crucial for subsequent studies that are based on spike analysis. Many spike sorting methods have been proposed, but universally adopted methods are not available yet. In this thesis, we introduce multilayer perceptrons (MLPs) that are trained, in unsupervised fashion, to minimize the Kullback-Leibler divergence (KLD) between original data and low-dimensional data probability distributions, thus leading to a low-dimensional data representation from which spike sorting problems are efficiently solved. More specifically, the proposed KLD-MLP algorithm learns a map from the original data space to a 2-D space where otherwise implicit spike clusters are revealed. For overall spike sorting performance comparison, four other algorithms have been applied at the data mapping stage: t-distributed stochastic neighbor embedding (t-SNE), principal component analysis (PCA), linear discriminant analysis (LDA), and locality preserving projections (LPP). The performance comparison is based on two publicly available synthetic datasets: the first one contains simulated spikes from a number of biological neurons ranging from two to twenty, and the second one contains simulated spikes from three neurons under different noise conditions. The KLD-MLP and t-SNE approaches yield significantly improved maps into the two-dimensional space where clustering is performed based on conventional K-means. The performance of basic K-means clustering based on KLD-MLP maps is maintained as the number of neurons or noise level are increased, which indicates that the method is potentially useful for spike analysis applications based on spike sorting.