Validation and Verification

This chapter discusses important aspects of the validation and verification of neural network models including selection of appropriate error metrics, analysis of residual errors and resampling methodologies for validation under conditions of sparse data. Error metrics reviewed include mean absolute error, root mean squared error, percent good and absolute distance error. The importance of interpretive schedules for pattern classification tasks is explained. The resampling methods bootstrap, jackknife and cross validation are compared to the popular neural network validation methods of train-and-test and resubstitution. While the techniques covered in this chapter are especially aimed at supervised networks where errors during training and validation are known, many of the techniques can be applied directly, or in a modified form, for unsupervised (self-organizing) neural networks.