CMHNN: a constructive modular hybrid neural network for classification

We propose a constructive RBF-like network that is able to learn discriminant functions in a multiclass classification problem where patterns are not individually labeled, but they belong to a higher level structure where knowledge about classes is present. The main differences with the standard RBF approaches can be summarized in two points. The number of localized receptive field (LRF) units is not fixed beforehand. Instead of it, we create a modular hidden layer with a constructive criteria that allows adding and updating units to each module. The supervised learning procedure doesn´t search for a minimum of the error function; it is a decision-based method that updates the connections from each hidden module to the output and affects the creation of LRF units. This architecture has rendered very good results on the classification of real images drawn from the database created for the ALINSPEC project