Initialization and construction of locally tuneable neural networks

This paper is a report of the work done on initialization and construction of locally tunable artificial neural networks using prototypes. This work started with the introduction of the interpolation networks (IN). These networks of locally tunable units are particularly well suited to a prototype-based initialization. Several experiments were conducted combining several types of networks (including IN) with competitive learning and prototype-based initializations. A method of construction of artificial neural networks, using prototypes, was also studied. The work also addressed the study a particular type of hybrid network that uses competitive learning to identify efficient initializations for new units in constructive algorithms.