A Two-Stage Subspace Trust Region Approach for Deep Neural Network Training

Abstract : In this paper, we develop a novel second-order method for training feed-forward neural nets. At each iteration, we construct a quadratic approximation to the cost function in a low-dimensional subspace. We minimize this approximation inside a trust region through a two-stage procedure: first inside the embedded positive curvature subspace, followed by a gradient descent step. This approach leads to a fast objective function decay, prevents convergence to saddle points, and alleviates the need for manually tuning parameters. We show the good performance of the proposed algorithm on benchmark datasets.
Type de document :
Communication dans un congrès
25th European Signal Processing Conference (EUSIPCO 2017), Aug 2017, Kos Island, Greece. pp.291-295, 2017, Proceedings of the 25th European Signal Processing Conference
Liste complète des métadonnées

Littérature citée [22 références]  Voir  Masquer  Télécharger

https://hal.archives-ouvertes.fr/hal-01634538
Contributeur : Emilie Chouzenoux <>
Soumis le : mardi 14 novembre 2017 - 11:32:35
Dernière modification le : jeudi 5 juillet 2018 - 14:45:35
Document(s) archivé(s) le : jeudi 15 février 2018 - 14:50:52

Fichier

PID4857015.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01634538, version 1

Citation

Viacheslav Dudar, Giovanni Chierchia, Emilie Chouzenoux, Jean-Christophe Pesquet, Vladimir Semenov. A Two-Stage Subspace Trust Region Approach for Deep Neural Network Training. 25th European Signal Processing Conference (EUSIPCO 2017), Aug 2017, Kos Island, Greece. pp.291-295, 2017, Proceedings of the 25th European Signal Processing Conference. 〈hal-01634538〉

Partager

Métriques

Consultations de la notice

331

Téléchargements de fichiers

104