Name
Abstract | ||
|---|---|---|
We propose a novel optimization algorithm which overcomes two drawbacks of Amari's natural gradient updates for information geometry. First, prewhitening the tangent vectors locally converts a Riemannian manifold to an Euclidean space so that the additive parameter update sequence approximates geodesics. Second, we prove that dimensionality reduction of natural gradients is necessary for learning multidimensional linear transformations. Removal of minor components also leads to noise reduction and better computational efficiency. The proposed method demonstrates faster and more robust convergence in the simulations on recovering a Gaussian mixture of artificial data and on discriminative learning of ionosphere data. |
Year | DOI | Venue |
|---|---|---|
2007 | 10.1109/IJCNN.2007.4371149 | IJCNN |
Keywords | Field | DocType |
amari natural gradient,optimisation,differential geometry,principal natural gradients,multidimensional linear transformations,euclidean space,learning (artificial intelligence),gaussian mixture,dimensionality reduction,discriminative learning,riemannian manifold,optimization,gradient methods,gaussian processes,tangent vectors,geodesic update,information geometry,ionosphere,linear transformation,noise reduction,discrimination learning,learning artificial intelligence | Information geometry,Dimensionality reduction,Computer science,Riemannian manifold,Tangent vector,Euclidean space,Differential geometry,Gaussian process,Artificial intelligence,Machine learning,Geodesic | Conference |
ISSN | ISBN | Citations |
1098-7576 E-ISBN : 978-1-4244-1380-5 | 978-1-4244-1380-5 | 1 |
PageRank | References | Authors |
0.39 | 0 | 2 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Zhirong Yang | 1 | 289 | 17.27 |
| Jorma Laaksonen | 2 | 1162 | 176.93 |