Autoencoder deep neural networks and their applications in solid-state physics
Low Temperature Physics · 01.06.2026 · 52 (6) · 660-666
DOI: 10.1063/10.0044110 ↗Kurz gefasst
Fachartikel in der Zeitschrift Low Temperature Physics, erschienen am 01.06.2026. Verfasst von V. Slavin. Die Arbeit nennt 26 Literaturangaben.
Abstract (Originalsprache)
We investigate the spectral properties of input–output Jacobians arising in deep neural network (DNN) autoencoders. These Jacobians are crucial in various applications of DNNs in solid-state physics, for instance as tools for detecting phase transitions. In this work, we study untrained DNNs with random, light-tailed initializations of weight matrices and bias vectors. We derive an explicit expression for the maximal singular value of the Jacobian for an arbitrary number of DNN layers, L. The maximal singular value is an important characteristic, as it determine the contraction properties and fixed points of autoencoders. At the same time, a number of phase transitions in solid-state physics can be described in terms of these fixed points. In the limit of large L, we obtain the equation governing the distribution of singular values of these Jacobians and provide its exact solution.
Zitieren
APA
Slavin, V. (2026). Autoencoder deep neural networks and their applications in solid-state physics. Low Temperature Physics, 52(6), 660-666. https://doi.org/10.1063/10.0044110
Vancouver
Slavin V. Autoencoder deep neural networks and their applications in solid-state physics. Low Temperature Physics. 2026;52(6):660-666. doi:10.1063/10.0044110
BibTeX
@article{Slavin2026,
author = {Slavin, V.},
title = {Autoencoder deep neural networks and their applications in solid-state physics},
journal = {Low Temperature Physics},
year = {2026},
volume = {52},
number = {6},
pages = {660--666},
doi = {10.1063/10.0044110}
}
EUVIMED