A Bayesian Account for Estimating the Number of Neurons during Spike Sorting
Keywords
No Thumbnail Available
Authors
Issue Date
2018-08-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Extracellular recordings have long been an invaluable tool for understanding neural
population activity. Spike sorting is the process of unmixing the contributing sources in
a recording to obtain the spiking activity of individual neurons. Identifying the correct
number of neurons is an error-prone process involving a considerable amount of intrinsic
uncertainty. However, most spike sorting algorithms do not account for this uncertainty,
but instead use a single point estimate. Using a fully probabilistic approach, we
demonstrate that the point estimate leads to systematic misestimation of the number of
neurons.
We estimate the number of neurons present in the data by sampling from the actual
posterior distribution using reversible jump Markov chain Monte Carlo, in the context
of realistic ground truth data. The expected value of the probabilistic estimate is then
compared to the widely used maximum a posteriori (MAP) estimate of the number of
neurons.
We find that even in the absence of incorrect modelling assumptions, using a point
estimate leads to a systematic underestimation of the number of present neurons. This
effect is visible for a wide range of values for the recording time and the noise available
in the recording. More specifically, we find that decreasing noise leads to a decrease in
this bias only for high sorting accuracy. If the sorting accuracy is low, this effect is
reversed. Furthermore, we find that the size of the bias can initially be decreased by
increasing the recording time, but for longer recordings this effect comes to a halt.
Misestimating the number of neurons contributes to errors in dividing spikes into
clusters, and thus impacts the clarity of the results, e.g. by fusing different neurons, or
splitting single neurons. As a consequence, correlations and other estimated properties
would be affected. The present results provide an analytical guide to correct for this
error.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen