Dynamics of representation dimensionality and class information during training of a spiking neural network
Received: 09 Oct 2026 Revised: 16 Oct 2026 Accepted: 08 Nov 2026
Published: 2026, vol. 30, issue 4, pp. 393–396
Abstract
For a single spiking ResNet-18 trained on image classification, we track, in all 17 layers of spiking neurons during training, the dimensionality of representations and the class information carried by individual neurons. In deep layers the dimensionality first drops and only then grows, and a linear measure does not detect this compression. Per-neuron class information rises in early and middle layers during the first epochs, then declines there and rises in deep layers; the epoch of its maximum does not decrease with layer depth.
Keywords: spiking neural networks, intrinsic dimension, mutual information, population coding.
BibTeX
@article{IS-Pospelov-Plusnin-Rogozhnikova-Ivanova-Toropova-Ivashkina-Avetisov-Anokhin2026,
author = {Pospelov, Nikita Andreevich and Plusnin, Viktor Vladimirovich and Rogozhnikova, Olga Sergeevna and Ivanova, Anna Andreevna and Toropova, Ksenia Aleksandrovna and Ivashkina, Olga Igorevna and Avetisov, Vladik Avanesovich and Anokhin, Konstantin Vladimirovich},
title = {{Dynamics of representation dimensionality and class information during training of a spiking neural network}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {393--396},
}
AMSBIB
\Bibitem{IS-Pospelov-Plusnin-Rogozhnikova-Ivanova-Toropova-Ivashkina-Avetisov-Anokhin2026}
\by N.\,A.~Pospelov, V.\,V.~Plusnin, O.\,S.~Rogozhnikova, A.\,A.~Ivanova, K.\,A.~Toropova, O.\,I.~Ivashkina, V.\,A.~Avetisov, K.\,V.~Anokhin
\paper Dynamics of representation dimensionality and class information during training of a spiking neural network
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 393--396
\lang In Russian
RU