Compact representation of position in autoencoders of neural activity
Received: 09 Nov 2026 Revised: 10 Nov 2026 Accepted: 10 Nov 2026
Published: 2026, vol. 30, issue 4, pp. 522–525
Abstract
Neural population activity is described by low-dimensional codes, and the method decides what can be read from the code. On synthetic calcium imaging recordings we compared an autoencoder with linear and manifold learning methods. When cells encode only the animal’s position, UMAP conveys it best; when the activity carries other signals too, position is read most accurately from the autoencoder’s code by a linear decoder at 8 dimensions, and when cells encoding a signal unrelated to behavior outnumber place cells, position is predicted more accurately from its 5-dimensional code than from the codes of the other methods.
Keywords: autoencoder, dimensionality reduction, place cells, synthetic data.
BibTeX
@article{IS-Orobets-Pospelov-Plusnin-Anokhin2026,
author = {Orobets, Margarita Alekseevna and Pospelov, Nikita Andreevich and Plusnin, Viktor Vladimirovich and Anokhin, Konstantin Vladimirovich},
title = {{Compact representation of position in autoencoders of neural activity}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {522--525},
}
AMSBIB
\Bibitem{IS-Orobets-Pospelov-Plusnin-Anokhin2026}
\by M.\,A.~Orobets, N.\,A.~Pospelov, V.\,V.~Plusnin, K.\,V.~Anokhin
\paper Compact representation of position in autoencoders of neural activity
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 522--525
\lang In Russian
RU