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. |
Year | DOI | Venue |
---|---|---|
2018 | 10.23919/EUSIPCO.2017.8081215 | 2017 25th European Signal Processing Conference (EUSIPCO) |
Keywords | DocType | Volume |
Deep learning,second-order approach,noncon-vex optimization,trust region,subspace method | Journal | abs/1805.09430 |
ISSN | ISBN | Citations |
2076-1465 | 978-1-5386-0751-0 | 0 |
PageRank | References | Authors |
0.34 | 14 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Viacheslav Dudar | 1 | 0 | 0.34 |
Giovanni Chierchia | 2 | 176 | 14.74 |
Emilie Chouzenoux | 3 | 202 | 26.37 |
Jean-Christophe Pesquet | 4 | 18 | 11.52 |
Vladimir Semenov | 5 | 0 | 0.34 |