Self-Optimizing Neural Network 3

Abstract

This paper describes an efficient construction of a partially-connected multilayer architecture and a computation of weight parameters of Self-Optimizing Neural Network 3 (SONN-3) that can be used as a universal classifier for various real, integer or binary input data, even for highly non-separable data. The SONN-3 consists of three types of neurons that play an important role in a process of extraction and transformation of important features of input data in order to achieve correct classification results. This method is able to collect and to appropriately reinforce values of the most important input features so that achieved generalization results can compete with results achieved by other existing classification methods. The most important aspect of this method is that it neither loses nor rounds off any important values of input features during this computation and propagation of partial results through a neural network, so the computed classification results are very exact and accurate. All the most important features and their most distinguishing ranges of values are effectively compressed and transformed into an appropriate network architecture with weight values. The automatic construction process of this method and all optimization algorithms are described here in detail. Classification and generalization results are compared by means of some examples.