Analysis of natural gradient descent for multilayer neural networks

Natural gradient descent is a principled method for adapting the parameters of a statistical model on-line using an underlying Riemannian parameter space to redefine the direction of steepest descent. The algorithm is examined via methods of statistical physics that accurately characterize both transient and asymptotic behavior. A solution of the learning dynamics is obtained for the case of multilayer neural network training in the limit of large input dimension. We find that natural gradient learning leads to optimal asymptotic performance and outperforms gradient descent in the transient, significantly shortening or even removing plateaus in the transient generalization performance that typically hamper gradient descent training.