A hessian matrix approach for training nonlinear networks
Changhua Yu, MICHAEL T. MANRY · 2005
In the original output weight optimization-hidden weight optimization (OWO-HWO) algorithm for training multilayer perceptions, only first order information is used to construct the desired net function. This gradient-like strategy inevitably reduces efficiency. In this paper, an efficient Hessian matrix inversion method is proposed for the hidden weights optimization. Numerical results validate the improvement of this algorithm.