Abstract
Aritificial Neural Network (NN) is applied to control, recognition and so on. The multi-layered Neural Network with sigmoid function (NNS) is often used in these fields because the NNS has two abilities of the interpolation and generalization. The generalization is, for example, for NN to recognize unlearned patterns to a certain extent. After this, we discuss recognition problems using NN. The NNS must learn both unlearned patterns and patterns given before to memorize unlearned patterns additionally, because the NNS cannot learn the patterns additionally. ART (Adaptive Resonance Theory) model can memorize the patterns additionally. However the ART model cannot classify patterns which have same pattern vectors.
A new neuron model called a Neural Network based on the distance between patterns (NDP) is proposed in this paper. The NDP has similar response functions to the Radial Basis Function. The NDP learns patterns by varying regions of neurons with the BP algorithm and adding new neurons. It is shown from experimental results on image recognition that the NDP can memorize patterns additionally and recognize unlearned patterns to some degree.
