Communication Dans Un Congrès Année : 2024

On divergence-free neural ODE for classification

Résumé

Inspired by the principles of fluid dynamics, divergence-free neural ODEs represent a promising class of neural networks that can improve classification performance and robustness. In this work, we propose methods for constructing divergence-free vector fields in high-dimensional latent spaces and thus extending their application to real-world scenarios. Our experimental results show that divergence-free neural ODEs can outperform their conventional counterparts in three classification tasks. In particular, divergence-free models show remarkable resilience to adversarial perturbations. Furthermore, divergence-free models are more effective at preserving the underlying structure of the data. This research provides a physics-based perspective on machine learning applications, in particular classification tasks, by interpreting neural networks as dynamical systems.

Fichier principal
Vignette du fichier
ICPR-2.pdf (1.2 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05120232 , version 1 (19-06-2025)

Licence

Identifiants

Citer

Zakaria Jarraya, Lucas Drumetz, Simon Benaïchouche, Douraied Ben Salem, François Rousseau. On divergence-free neural ODE for classification. ICPR 2024: 27th International Conference on Pattern Recognition, Dec 2024, Kolkata, India. pp.254-265, ⟨10.1007/978-3-031-78186-5_17⟩. ⟨hal-05120232⟩
124 Consultations
218 Téléchargements

Altmetric

Partager

  • More