A Dynamical Systems Analysis of Sequential Hopfield Networks With Softmax Activation
Neural Computation · 05.11.2026 · 1-18
DOI: 10.1162/neco.a.1603 ↗Kurz gefasst
Fachartikel in der Zeitschrift Neural Computation, online erschienen am 05.11.2026. Verfasst von Sandro Zampieri.
Abstract (Originalsprache)
Classical associative memory models describe retrieval as a static relaxation process toward stable attractors, failing to autonomously generate sequences of states. While recent architectures overcome this limitation by introducing slowly evolving auxiliary variables to trigger transitions between memories, their mathematical analysis has primarily focused on storage capacity rather than explicit dynamical behavior. In this work, we develop a rigorous analytical framework to investigate the dynamics of a sequential Hopfield-type network coupled with a linear slow variable under a low-temperature softmax activation function. By focusing on orthogonal memories, we leverage geometric singular perturbation theory and fixed-point analysis to achieve a major simplification of the model, enabling the precise characterization of oscillation amplitudes, escape times, and other key dynamical features. Numerical experiments demonstrate that our findings remain highly informative for the standard random memories setup.
Zitieren
APA
Zampieri, S. (2026). A Dynamical Systems Analysis of Sequential Hopfield Networks With Softmax Activation. Neural Computation, 1-18. https://doi.org/10.1162/neco.a.1603
Vancouver
Zampieri S. A Dynamical Systems Analysis of Sequential Hopfield Networks With Softmax Activation. Neural Computation. 2026:1-18. doi:10.1162/neco.a.1603
BibTeX
@article{Zampieri2026,
author = {Zampieri, Sandro},
title = {A Dynamical Systems Analysis of Sequential Hopfield Networks With Softmax Activation},
journal = {Neural Computation},
year = {2026},
pages = {1--18},
doi = {10.1162/neco.a.1603}
}
EUVIMED