A Neuron Coverage-Based Self-organizing Approach for RBFNNs in Multi-class Classification Tasks

Radial Basis Function Neural Networks (RBFNN) constitute a distinct type of artificial neural network known for its unique architecture and ability to approximate complex functions, and as well lead to effective modeling in regression/classification tasks. Within the context of multi-class classification, this work introduces NC-SORBFNN as a self-organizing approach that adapts the structure of RBFNNs to the complexity of each task. NC-SORBFNN comprises an initialization step that makes use of a Self-Organizing Map (SOM) to give rise to a first set of hidden neurons that is next tuned by a self-adjustment process that goes through the data to make the network grow and shrink as required on the basis of the sample coverage of each neuron. The experimental results show the effectiveness of NC-SORBFNN, which in turn compares well with other similar solutions, leading in general to high performance for a low amount of hidden neurons.