English

Accelerating Natural Gradient with Higher-Order Invariance

Machine Learning 2018-06-11 v2

Abstract

An appealing property of the natural gradient is that it is invariant to arbitrary differentiable reparameterizations of the model. However, this invariance property requires infinitesimal steps and is lost in practical implementations with small but finite step sizes. In this paper, we study invariance properties from a combined perspective of Riemannian geometry and numerical differential equation solving. We define the order of invariance of a numerical method to be its convergence order to an invariant solution. We propose to use higher-order integrators and geodesic corrections to obtain more invariant optimization trajectories. We prove the numerical convergence properties of geodesic corrected updates and show that they can be as computationally efficient as plain natural gradient. Experimentally, we demonstrate that invariance leads to faster optimization and our techniques improve on traditional natural gradient in deep neural network training and natural policy gradient for reinforcement learning.

Keywords

Cite

@article{arxiv.1803.01273,
  title  = {Accelerating Natural Gradient with Higher-Order Invariance},
  author = {Yang Song and Jiaming Song and Stefano Ermon},
  journal= {arXiv preprint arXiv:1803.01273},
  year   = {2018}
}

Comments

ICML 2018

R2 v1 2026-06-23T00:41:09.452Z