Effects of training data distribution on backpropagation generalization

The authors investigate the effects of the training data set distribution on the generalization capability of multilayer neural networks with backpropagation. The sensitivity of the neural network performance to the training data set distribution and how this performance varies are shown visually. An appropriate way to select this distribution is outlined. The results are presented by using a medical diagnosis pattern recognition problem, with different diagnosis of one main symptom: vaginal discharge of nonmenstruating women. This database permits easy generation of typical data of simulated patients by Monte Carlo random simulation.<
>