EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Finite-Time Lyapunov Exponents of Deep Neural Networks

L Storm, H Linander, J Bec, K Gustavsson, B Mehlig

Physical Review Letters · 01.02.2024 · 132 (5)

DOI: 10.1103/physrevlett.132.057301 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physical Review Letters, online erschienen am 01.02.2024. Verfasst von L Storm und 4 weiteren Autor:innen. Beteiligte Einrichtung: Department of Physics, University of Gothenburg, 41296 Gothenburg, Sweden. Die Arbeit wurde bisher 22-mal zitiert (Crossref) und nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Knut och Alice Wallenbergs Stiftelse, Vetenskapsrådet, Université Côte d’Azur.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We compute how small input perturbations affect the output of deep neural networks, exploring an analogy between deep feed-forward networks and dynamical systems, where the growth or decay of local perturbations is characterized by finite-time Lyapunov exponents. We show that the maximal exponent forms geometrical structures in input space, akin to coherent structures in dynamical systems. Ridges of large positive exponents divide input space into different regions that the network associates with different classes. These ridges visualize the geometry that deep networks construct in input space, shedding light on the fundamental mechanisms underlying their learning capabilities. Published by the American Physical Society 2024

Autor:innen und Institutionen

  • Department of Physics, University of Gothenburg, 41296 Gothenburg, Sweden.
  • Department of Mathematical Sciences, Chalmers Technical University and University of Gothenburg, Gothenburg, Sweden.
  • MINES Paris, PSL Research University, CNRS, Cemef, Valbonne, F-06900, France.
  • Université Côte d'Azur, Inria, CNRS, Cemef, Valbonne, F-06900, France.

Förderung

  • Knut och Alice Wallenbergs Stiftelse
  • Vetenskapsrådet
  • Université Côte d’Azur

Zitieren

APA

Storm, L., Linander, H., Bec, J., Gustavsson, K., & Mehlig, B. (2024). Finite-Time Lyapunov Exponents of Deep Neural Networks. Physical Review Letters, 132(5). https://doi.org/10.1103/physrevlett.132.057301

Vancouver

Storm L, Linander H, Bec J, Gustavsson K, Mehlig B. Finite-Time Lyapunov Exponents of Deep Neural Networks. Physical Review Letters. 2024;132(5). doi:10.1103/physrevlett.132.057301

BibTeX
@article{Storm2024,
  author = {Storm, L and Linander, H and Bec, J and Gustavsson, K and Mehlig, B},
  title = {Finite-Time Lyapunov Exponents of Deep Neural Networks},
  journal = {Physical Review Letters},
  year = {2024},
  volume = {132},
  number = {5},
  doi = {10.1103/physrevlett.132.057301}
}
Kostenlos starten →