Localized generalization error model for Multilayer Perceptron Neural Networks

In this work, the localized generalization error model (L-GEM) for Multilayer Perceptron Neural Network (MLPNN) is derived. The L-GEM is inspired by the fact that a classifier should not be required to recognize unseen samples that are very different from the training samples. Therefore, evaluating a classifier by very different unseen samples may be counter-productive. In the L-GEM, the "local" is defined by the difference between feature values of unseen samples and training samples is less than a given real value (Q). The L-GEM provides an upper bound of the Mean-Square-Error of unseen samples "local" to the training dataset. As the generalization capability of a MLPNN is the key evaluation criterion of a successful training of MLPNN, we select the number of hidden neurons of a MLPNN using the L-GEM. The experimental results on four UCI datasets show that the proposed L-GEM yields better MLPNNs with higher generalization power (testing accuracy) and smaller number of hidden neurons.