Non-linear dimensionality reduction techniques for spike sorting

EMILIO KROPFF; JOAQUIN SEIA; JORGE SANCHEZ

Congreso; Reunion anual de la Sociedad Argentina de Neurociencias; 2019

Spike sorting is a key step in the processing of extracellular electrophysiologicaldata, in which recorded spikes are clustered based on their shape, ideally reflectingthe different neurons that originated them. The improvement in the acquisitionhardware has allowed an exponential growth in the number of neurons that can beregistered in parallel, but spike sorting algorithms have not advanced at a similarpace.Several steps need to be followed in order to isolate and classify recorded spikes.After preprocessing the signal and detecting the spikes, a feature extraction anddimensionality reduction process takes place. Finally, the low dimension featurevectors are clustered manually or automatically.In this work, we present a comparative study between two non-lineardimensionality reduction algorithms that could improve manual and automaticsorting: T-Distributed Stochastic Neighbor Embedding (t-SNE)[1] and UniformManifold Approximation and Projection (UMAP)[2].