A rapid multi-layer perceptron training algorithm

A fast algorithm for training multilayer perceptrons is presented as an alternative to the backpropagation algorithm. Training time is reduced considerably over standard backpropagation. The philosophy behind this method is the same as that introduced by R. Scalero and N. Tepedelenlioglu (1991), where the normal equation is obtained and solved at every node in the network using a Kalman filter. The algorithm introduced replaces the Kalman filter by a conjugate-gradient method of updating. This algorithm shortens the training time by several orders of magnitude for the classification problem considered