A Two-Stage Subspace Trust Region Approach for Deep Neural Network Training - Archive ouverte HAL Access content directly
Conference Papers Year : 2017

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.
Fichier principal
Vignette du fichier
PID4857015.pdf (310.45 Ko) Télécharger le fichier
Origin : Files produced by the author(s)
Loading...

Dates and versions

hal-01634538 , version 1 (14-11-2017)

Identifiers

  • HAL Id : hal-01634538 , version 1

Cite

Viacheslav Dudar, Giovanni Chierchia, Emilie Chouzenoux, Jean-Christophe Pesquet, Vladimir V. 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. ⟨hal-01634538⟩
358 View
296 Download

Share

Gmail Facebook X LinkedIn More