Structure training of neural networks

Previously, a procedure for designing multilayer feedforward neural networks has been advanced based on the construction of a Voronoi diagram (VOD) in multidimensional feature space. Here, the advantage of the approach in realizing the important property of robust generalization, which demands satisfactory performance in cases where an uncertain test input pattern deviates from an exemplar is analyzed and illustrated by application to the d-bit parity problem. Next, it is shown how a neural network may be obtained directly from the Delaunay tessellation which is the abstract dual of the Voronoi diagram.<<ETX>>