Nicol N. Schraudolph and Thore Graepel
The method of conjugate gradients provides a very effective way to optimize large, deterministic systems by gradient descent. In its standard form, however, it is not amenable to stochastic approximation of the gradient. Here we explore a number of ways to adopt ideas from conjugate gradient in the stochastic setting, using fast Hessian-vector products to obtain curvature information cheaply. In our benchmark experiments the resulting highly scalable algorithms converge about an order of magnitude faster than ordinary stochastic gradient descent.
|Published in||Proceedings of the 9th International Conference on Neural Information Processing, ICONIP 2002|