An algorithm for model determination in a layered network with non-uniform hidden layer unit set

On applying a layered neural network to pattern classification problems, simultaneous search for the optimal model and the optimal parameter set can be beneficial, in order to solve the model determination problem. When altering the model during training, it is desirable that the accumulated map will be inherited to the new model for the training process to be continued. In this work, a model selection criterion based on inter-map norms will be introduced, for mappreserving Model Switching applied to networks with non-uniform hidden layer unit sets. Further, a search algorithm for the suitable model applicable to problems requiring various types of discrimination borders, will be introduced.