Interpolation Networks

Abstract

This paper introduces a new type of network based on local response units, the ‘Intelpolation Networh’ (INS).U nder certain conditions this network is an interpolator. Its formulation allows a type of initialisation by prototypes that will set the net in a good initial starting point for the subsequent supervised learning process. These networks can be seen as a type of ‘Radial Basis Functions Network’ (RBFN) [Moody88]. However, their basis functions are essentially inverse squared distances instead of Gaussian functions. A brief description of the origin and motivation of INS is made in the first section, followed by the description of the first experiments with these networks.