• русский
• український
• english

Abstract

Classification using neural networks

Why exactly artificial neural network

Recently computers distribution in all places makes more and more complicated objectives for the last ones. But even increasing performance can't help computer to solve short formalized tasks. Neural networks exceed sequential computers in solving tasks, where human exceed a computer. Solving high accuracy and high quantity calculations tasks on usual computers is preferred. Neural networks used to solve next tasks:

• visual and audio recognition; area(field) of application: text recognition, target recognition on radar screen, voice-controlled systems;
• associative searching and associative model creation; speech synthesis; natural language forming;
• models and nonlinear hard-describable mathematical systems forming and evolution prognostication;
• using in a production; cyclones and other natural processes evolution prognostication; rate of exchange changing and other financial processes;
• control and regulation systems with forecast possibility; robots and other complex devices control;
• automatic machines: queuing and commutation systems, telecommunication systems;
• diagnostics and decision tacking, excluding logical conclusion: in medicine, in crime detection, in finances;

Neural networks unique kind is universality. For all this tasks exists effective mathematical solutions. Neural networks yields to particularized methods for concrete tasks, but owing to universality, they are important research direction, the subject of much study.

Rich potential

Neural networks are a powerful modeling method, which allows reproducing extremely complicated dependences. Particularly neural networks are nonlinear by their nature. During many years linear modeling was the basic modeling method in many areas of application, because optimization procedures were developed for it. In tasks where linear approximation is inadequate linear models works badly (if they works at all). Also, neural networks able to model linear dependences with a big variable quantity.

Using simplicity

Neural networks teaching on different examples. Neural network user selects representative data and starts learning algorithm, which is automatically apprehend data structure. User must give some set of heuristic knowledge for data selection, network architecture structure choice and data interpretation. But needed knowledge level for successfully neural network using is more modest than at traditional statistic methods.

Research tasks and objectives

Classification studying using neural networks supposes existing neural networks architectures and learning methods studying, choosing optimal architecture for solving the problem and experimental research. Finding new fields of application of this technology is a separate task of research.

Neural networks architectures

There is much different neural networks architecture known in nowadays. Every variant has its own advantages and disadvantages, realization and teaching features. Research aim is to select the most proper architecture types for solving the problem (classification problem is a composite problem which includes many sub-problems, for solving each of they used its own architecture type).

Neural network teaching

Neural network teaching putting into practice by three ways: teaching with a teacher help, teaching without teacher, combined method. The way of network teaching depended with networks architecture. Teaching excerpts values vagueness existing must also be taken into account.

Optimization

It is needed to develop artificial neural network components software model for the most proper architecture types revealing. Also needed to research different architecture teaching methods and make experimentally track architecture ability to adequate solving needed problem.

Scientific newness

Assumed scientific newness is in formation of rules, which can be useful in neural network based systems development. Nowadays this rules are not existing, only just some guidelines for neural networks using.

Also, scientific newness is in new application methods developing. The main direction of developing is research of potential abilities of neural networks to help to peoples with limited physical possibilities.