This paper proved that the multi-layered feed-forward neural network with linear output nodes can be used as an optimal feature extractor.It is also proved that the output function of a classifier network is a least-meansquare approximation to the Bayes decision function.For a threelayered network with linear output nodes

arbitrary approximation precision can ha obtained if the network has enough hidden nodes.On the basis of these conclusions

a combined netwrok with the functions of both feature extraction and decision is proposed.The result shown that this combined network has better properties than that of a single network.