Mapping Symbolic Knowledge into Locally Receptive Field Networks

This paper investigates Locally Receptive Field Networks, a broad class of neural networks including Probabilistic Neural Networks and Radial Basis Function Networks, which naturally exhibit symbolic properties. Moreover, specific attention is given to the sub-class of Factorizable Radial Basis Function Networks whose architecture can be directly translated into a propositional theory and viceversa. Exploiting this characteristics, symbolic and numeric algorithms can be easely integrated for automating network synthesis. Several methods including classification and regression trees, and statistical clustering are evaluated on a classification task in a difficult medical domain. The obtained results show that the considered network class is able to achieve a high accuracy, while conserving a symbolic readability.