re,
NEURON NETWORKS WITH NONLINEAR INTERCONNECTIONS FOR ANALYSIS OF DATA.
iral Ivanov A.N.
associated professor of Moscow State Technical University,
der Moscow, Russia, #158
Abstract - Structure of neuron networks with
ime different types of interconnections is described.
Analitical method, based on nonlinear spectral
na. representation of optical information is used to
classify real signals and objects corresponding to
them.The important idea of time evolution of
of spectral parameters is then introduced and it is
) shown that the problem of parallel signal processing
, can be solved with the aid of large number of
ing identical elements with nonlinear interconnections.
Problems, connected with distortion and signal
correction as well qualitatively discussed.It is
investigated influence of losses along network on
signal propagation.Main characteristics of nonlinear
neuron network were marked out.2 D neuron networks
with nonlinear interconnections for analysis of
datas were investigated as well.Models of real
neuronlike elements were suggested.Their regimes
were investigated, simulating neurons with the aid
of digital computer.
Experimental results enabled to make a
conclusion of significant possibilities which arise
when using networks with nonlinear interconnections.
It is marked some problems and applications inherent
to these networks.
KEY WORDS : Neuron, network, soliton, spectrum,
filtration.
1. INTRODUCT ION nature of these interconnections may be different.
It is a poor idea to imitate the behaviour of
In the systems of remote sensing investigator neurons. It’s more productive to use some principals
often meets with problems of preliminary signal inherent to neurons. Among them flexibility, access
processing.Having read signal from receiving matrix to a large amount of data and of course, opportunity
it is necessary to realize its transformation to to operate with digital and especially analog signal
form admissible for digital computer. Latter that most investigators find possible to neglect.
operates with high dimension matrix of datas using Almost all applications of neuron networks in
algorithms which depend on concrete task. Sometimes optical systems are based on ability to recognize
that sequence of actions is inconvenient. Time, standart image, perform some mathematical operati-
requested for computer, and registering processes ons. Whole neuron system is devided into several
velocity may come into contradiction. Besides, layers. They have interconnections which may be
distribution of operations between analog and called vertical, i.e. between different layers ( see
digital units is far from symmetry. And problem of Fig.la).In most situations each element of network
transference of processing center for whole system realizes addition and substraction. More complex
to analog unit arises. operations are performed by network as a whole. Thus
It has become possible to come to this a role of interconnections comes to a simple
conclusion after investigations connected with transportation of datas.
analysis of pictures through atmosphere. Discussed Imagine another situation when each neuron in
in /1/, acoustooptical device required preliminary the layer is connected with the neighbour one. Such
analog signal processing directly at the optical structure may be called as network with horizontal
matrix of photodetectors. Furthemore in perspective or parallel interconnections in layer (see Fig.1b).
it is necessary to organize parallel functioning of
all elements . That is why direction of N N N N 1
=
investigations was concentrated to biological | N | a va
"ii ln |
objects such as retina.It was determined that real N N N 2
N AN
neural system may be envisage as a model of complex EE
technical system for effective transformation of M N N N ! N 3
information. a) Linear networks
Human brain consists of a great number of unique
elements - neurons. Much time has been spent N=====N=====N=====N N =N===== N 1
investigating structure, interconnections and [ |
operations of them.It can be said that now we have N=====Nzz===zN====z=N====sN=====N=====N e
several lines of development those systems which are
often called neuron networks. Although the term b) Nonlinear networks
"neuron" may be taken as undefined it must be
restricted to avoid ambiquities. In this paper when Figure 1 Types of networks
we speak of "network" we should be refferring to
collections of physical elements. Their interconnec- Braph of all interconnections defines its topology.
tions provide all system with new functions. The Idea of symmetry plaies here an extraordinary role.
17
