Abstract

Neural networks based on RBFs have recently gained a wider acceptance over the backpropagation topology because of its faster learning rates due to the decoupling between the hidden and output layers.¹ We have derived still another network topology based on normalized RBFs, which do not require learning time. This normalization process makes the classifying network behave radically different from the conventional RBFs classifier, because it creates sharp classifying boundaries among the classes. This new topology has been applied to the classical problem of character recognition degraded by a noisy binary channel. The classifying properties obtained by this system are equivalent to a system formed by a classical RBF approximation followed by a winner-takes-all network (WTA), without the disadvantages of the convergence time required by a WTA network. Also, as mentioned above, a powerful advantage of this network is that is does not require learning time. Only a representative number of templates are needed; they can be fed into the network without any prior processing.

© 1991 Optical Society of America