Multilevel RBF to resolve classification problems with large training sets: new pseudo-exact procedure

This article describes a multilevel neural architecture based on SOMs and radial basis function networks, which can produce results and approximation of functions close to those obtained using a single radial basis function network. In addition, the learning algorithm of this new architecture delivers a very significant reduction in the training time of the network and, at the same time, allows parallelization techniques to be applied in a natural way. The proposed training system is not an iterative method, it is not exactly an exact one. Our proposal procedure requieres an amount of time similar to the iterative methods but with the efficiency of the exact methods. This architecture has been developed to test on problems involving ecological segmentation, as part of environmental study and monitoring programmes by research groups in southeastern Spain.