A non-iterative method for training feedforward networks

The authors describe a method for determining the weights of a feedforward network, with only one hidden layer, to perform a certain classification task. The Wiener least squares solution is part of this algorithm and is used to calculate weights for the input layer and the output units. This is a non-iterative method, and, from a computational point of view, it is much faster than standard methods. The algorithm presented is applied to three data sets with known statistical properties. The networks learn with relatively small data sets and the final networks are capable of classifying unknown patterns with classification errors approximating the Bayes error.<<ETX>>