нейронные сети

Algebraic properties of recurrent neural networks of discrete time

Authors: Slepovichev I. I.

Artificial neural networks can be used effectively for a quite general class of problems. Still there exists no formal foundation of some important constructions used in the theory. In this paper an attempt is undertaken to formalize some concepts of neuroinformatics and consider their properties from the point of view of applied universal algebra. It is proposed to treat neural networks as heterogeneous algebras which has made it possible to prove for them basic results similar to algebraic theorems on homomorphisms and congruences.

Learning Neural Network Controllers for Stabilizing Hybrid Dynamic Systems

Authors: Andreichenko D. K., Fedor M. Z.

Control modules based on artificial neural networks (NN) are often used for controlling objects with lumped parameters. Controled objects in such systems have finite set of natural oscillation frequencies.

The Application of Artificial Neural Networks to Identification of Some Amino Acids in Binary Mixtures

Authors: Aravin O. I., Novikov A. Y., Selifonova E. I., Chernova R. K., Shevirev S. P.

The paper develops the technique of applying the method of artificial neural networks for processing of the spectrophotometric data in order to determine the phenylalanine and tyrosine in undivided binary mixtures of these amino acids at microgramm concentrations. Calculated error in the determination is: minimum of 1%, the maximum does not exceed 10%. The maximum error is observed for mixtures in which the content components differ by an order or more.

Construction of Diagnostic Expert Systems on the Basis of Neural Networks

Authors: Solovyev V. M.

In work questions of construction of diagnostic expert systems (DES) on the basis of neural networks (NN) with lateral braking are considered. Methods of training of such networks are offered. Questions of reception of the diagnostic information in the heterogeneous computer network and uses of the aprioristic information on the importance of diagnostic attributes are analyses. Results of work can be used at construction NN diagnostic systems (clustering).

Algebraic Properties of Abstract Neural Network

Authors: Slepovichev I. I.

The modern level of neuroinformatics allows to use artificial neural networks for the solution of various applied problems. However many neural network methods put into practice have no strict formal mathematical substantiation, being heuristic algorithms. It imposes certain restrictions on development of neural network methods of the solution of problems. At the same time there is a wide class of mathematical models which are well studied within such disciplines as theory of abstract algebras, graph theory, automata theory.

Simultaneous Approximation of Polynomial Functions and Its Derivatives by Feedforward Artificial Neural Networks with One Hidden Layer N. S.

Authors: Uzentsova N. S., Sidorov S. P.

 In this paper we propose the algorithm for finding weights of feedforward artificial neural networks with one hidden layer to approximate polynomial functions and its derivatives with a given error. We use the rational sigmoidal function as a transfer function.