A new supervised training algorithm for generalised learning

Abstract

The paper proposes a new supervised training algorithm for feedforward neural networks. Instead of applying single valued input-output information, multivalued information in the form of a K-dimensional vector (K>1) is applied to each node of the input-output layer. Weights are adjusted using the gradient decent approximation method in order to minimise the sum-squared error value at each node of the output layer. The training algorithm has been studied for wide range of input-output values and gives worthy results especially when the output vector is small enough compared to the input vector. The paper suggests a judicious method for choosing the bias component of the sigmoidal activation function used in the training algorithm.