The paper introduces new valuable improvements of performance, a construction and a topology optimization of the Self-Optimizing Neural Networks (SONNs). In contrast to the previous version (SONN-2), the described SONN-3 integrates the very effective solutions used in the SONN-2 with the very effective ADFA algorithms for an automatic conversion of real inputs into binary inputs. The SONN-3 is a fully constructive ontogenic neural network classificator based on a sophisticated training data analysis that quickly estimates values of individual real, integer or binary input features. This method carries out all computation fully automatically from a data analysis and a data input dimension reduction to a computation of a neural network topology and its weight parameters. Moreover, the SONN-3 computational cost is equal O(n log 2 n), where n is a sum of a data quantity, a data input dimension and a data output dimension. The results of the SONN-3 construction and optimization are illustrated and compared by means of some examples.