A NEW KIND OF NEURAL NETWORK BASED ON RADIAL BASIS FUNCTIONS

We have derived a new kind of neural network using normalized radial b asis functions, ''RBF'', with the same classifying properties as if it were built up using sigmoid functions. This equivalence is mathematic ally demonstrated. In addition, to this, we also show that the propose d network is equivalent to a gaussian classifier. The network does not require any computing learning time to build a classifier. This netwo rk has been compared with well known adaptive networks, such as backpr opagation and linear combination of generalized radial basis functions (GRBF's). Its adapted forms are presented to see how the classifying regions and boundaries among the supplied examples are formed. This ne ural network can be made to have identical classifying properties, as the nearest neighborhood classifier ''NNC''. In the case of having man y examples per class, fewer centers can be found using vector quantizi ng ''VQ'' techniques as done in Kohonen's network. Finally, this neura l system can also be used to approximate a smooth continuous function, given sparse examples.