Multilayer feed-forward artificial neural networks for class-modeling

A class-modeling algorithm based on multilayer feed-forward artificial neural networks is proposed. According to this method, each category model is described by an auto-associator network, so the class space is defined on the basis of a distance to the model criterion which takes into account the residual standard deviation of the reconstructed input vectors. The details of the method are discussed and examples of its application to a simulated ("exclusive-OR") and a real-world (classification of wines) problem are presented. As far as the simulated highly non-linear example is concerned, NN-based class modeling outperforms SIMCA and UNEQ both in terms of classification rate and specificity. On the other hand, when dealing with the wine data set, which has a less non-linear structure, our proposed method still provides comparable and, in some cases, better results than the other two techniques.