hide
Free keywords:
prediction reliability; activation pattern; self-organising maps; radial basis function networks; hybrid training; visualisation techniques
Abstract:
A new approach to neural network models is able to overcome the black-box-problem of neural networks by producing a measure, how sure the network is about its answer. The principle idea behind this measure is to use a two-segmented network, where the first segment works as an input-oriented, (mostly trained by unsupervised methods) classification device, whereas the second segment produces the output based on the classification given by the first segment. An analysis how good an input fits into the given classification produces the measure for the quality of the network response. This measure is of course by no means of the quality of error bars produced by statistical methods, however it is a good indication of how close the given input is to those used for the training of the neural network. Neural network models with this two-segmented architecture are not new, (e.g. RBFN or counterpropagation networks), however they have not been used so far to obtain information about possible errors of the network. We apply this network to data on bioindication and niche identification of 10 small rivers in German low-mountains and a long-term study of a small stream in Schleiter et al., (this journal) and Obach et al., (this journal) and compare the results to different approaches. Conclusion: The new network approach is suitable for creating models that are capable to estimate the accuracy of their response even in the situation where only few data for training are available.
