Knowledge-increasable artificial neural network and natural gradient algorithm
Yaping Huang, Siwei Luo, Jian-Yu Li · 2002
Provides a knowledge-increasable artificial neural network model and learns parameters by using a probability model. A conventional gradient algorithm is normally adopted to learn parameters in a KI network, but its performance isn't the best. The paper uses the natural gradient algorithm that takes the Riemannian metric of parameter space to define the parameters. This method can adaptively modify the parameters based on a Riemannian metric and achieve the approximate best performance.