A new approach for dynamic node creation in multilayer neuralnetworks

摘要

An approach to simultaneous recursive weight adaptation and nodecreation in multilayer perceptron neural networks is presented. Themethod uses time and order update formulations in the orthogonalprojection method to arrive at a recursive weight updating procedure forthe training process of the neural network and a recursive node creationalgorithm for weight adjustment of a layer with added nodes during thetraining process. The approach allows optimal dynamic node creation inthe sense that the mean-squared error is minimized for each newtopology. The effectiveness of the algorithm was demonstrated on a realworld application for detecting and classifying underground dielectricanomalies