Instant learning for supervised learning neural networks: a rank-expansion algorithm
C. L. Philip Chen, Jiyang Luo · 2002
An one-hidden layer neural network architecture is presented. An instant learning algorithm is given to decide the weights of a supervised learning neural network. For an n dimensional, N-pattern training set, a maximum of N-r hidden nodes are required to learn all the patterns within a given precision (where r is the rank, usually the dimension, of the input patterns). Using the inverse of activation function, the algorithms transfer the output to the hidden layer, add bias nodes to the input, expand the rank of input dimension. The proposed architecture and algorithm can obtain either exact solution or minimum least square error of the inverse activation of the output. The learning error only occurs when applying the inverse of activation function. Usually, this can be controlled by the given precision. Several examples show the very promising result.>