Scatter-partitioning Rbf Network for Function Regression and Image Segmentation: Preliminary Results

Scatter-partitioning Radial Basis Function (RBF) networks increase their number of degrees of freedom with the complexity of an input-output mapping to be estimated on the basis of a supervised training data set. Due to its superior expressive power a scatter-partitioning Gaussian RBF (GRBF) model, termed Supervised Growing Neural Gas (SGNG), is selected from the literature. SGNG employs a one-stage error-driven learning strategy and is capable of generating and removing both hidden units and synaptic connections. A slightly modiied SGNG version is tested as a function estimator when the training surface to be tted is an image, i.e., a 2-D signal whose size is nite. The relationship between the generation, by the learning system, of disjointed maps of hidden units and the presence, in the image, of pictorially homogeneous subsets (segments) is investigated. Unfortunately, the examined SGNG version performs poorly both as function estimator and image segmenter. This may be due to an intrinsic inadequacy of the one-stage error-driven learning strategy to adjust structural parameters and output weights simultaneously but consistently. In the framework of RBF networks, further studies should investigate the combination of two-stage error-driven learning strategies with synapse generation and removal criteria. y Internal report of the paper entitled \Image segmentation with scatter-partitioning RBF networks: A feasibility study," to be presented at the conference