qualitatively to show opportunity of signal 
processing with the aid of neuron networks using non 
linear interconnections.Because of it technical 
realization of elements and their interaction was 
rather different from one discussed above.It was 
made so for more convenience to set initial 
conditions and to obtain results. 
Network, based on long line, was assembled 
making use of analog computer. Nonlinearities there 
were huild up using internal units inherent to these 
computers.Thus simulations were carried out with 12 
elements in network.It is convenient to form 
necessary initial charges at all knots.A schematic 
diagram of experimental system is shown in Fig.6. 
Output of each neuron has been taken from 
corresponding knot.Received sygnals moved to acousto 
-optical cell which had 6 parallel channels through 
frequency transformers.System of transformers could 
form 36 different channels (6 at one acousto-optical 
element).Each channel has own attenuator.In that way 
it was possible to regulate weights of outputs. 
Collection of moving plates had functional 
transparency along its lengh.All optical elements 
provide necessary characteristics of light beams at 
all points.Group of photodetectors registered 
signals which then were transformed to digital code. 
This experimental system enabled to realize discrete 
spectral filtration.As an initial image was chosen 
a rectangular with distorced plane top 
(simmetrically with respect to plane top).Moving 
different functionally transparent plates, and 
treating signals at outputs, were obtained two 
different results.First signal is undistorted 
rectangular with suppressed oscilations of plane top 
and second one is high frequency component with 
suppressed initial rectangular. 
Time control enabled to fix time of back 
calculation but it was considered only as first 
approximation and the finish result was fixed, using 
qualitative determination of expected signal.Real 
network and signal processing system may be formed 
on the base of VLSI structures, which will provide 
opportunities of operating with complicated images 
using many neuronlike elements. 
5. NETWORKS OF HIGH DIMENSION 
Previous discussion was concerned only linear 
networks.Nevertheless networks of more than first 
dimension can be devided into two classes. First is 
n-Dimensional network formed by only nonlinear 
interconnections. There are mathematical literature 
where 2D solitons are investigated.They are called 
lumps.Lumps are solitons of Kadomtsev-Petviashvili 
equation.It is a problem to find a suitable model of 
this mathematical ob ject.Spectral theory of 2D 
networks with non linear interconnections is not 
enough clear to build on the base of it system for 
signal processing.However principle of scattering in 
that theory can be used. 
Stream of information through neuron oscilations 
interacts with the set of centers scattering part of 
this information and producing two components of 
signal spectrum.Representation of 2D signals as a 
series and classification of them has the same 
application as for linear networks.It should be 
stressed that task of scattering acquires thus a new 
character.Using idea of absence of continuous 
spectrum component one can obtain instead of real 
image system of moving lumps.One can suppose that 
parameters of lumps have as well one-to-one 
correspondence with the Fourier spectrum of initial 
image. Then nonlinear spectral filtration would 
enable to pick out necessary elements from 
real image.If dimension of network is more than two, 
then it is difficalt to find fruitful concept of 
23 
interaction of neurons. Toda (/12/) has represented 
several matrices corresponding to differential 
equations of more high order than Korteveg-de-Vries 
one. Exact technical realization of these mathemati- 
cal objects is impossible, but may be approximation 
would enable to build new interesting network. 
The second type of 2 D network can be built 
organizing links between several ordinal networks. 
This problem is closely depends on the set of neuron 
parameters, introduced while describing scattering 
process. Fig.2 gives picture of distribution of 
several signals, directed to differrent network 
sections. Control solitons are formed not at all 
neurons. Two neurons, having the same scattering 
properties at © = 0, can differ at any different 
angles. In polar coordinates angle corresponds to 
solitons phase or its position at network. Radius 
defines normalized amplitude of soliton. However it 
is a special question. 
6. CONCLUSION 
The most important conclusion of this study is 
that it is possible to use neuron network with non- 
linear interconnections for signal processing. Ordi- 
nal vertical networks usually play role of some 
mathematical model and technical realization by 
digital computer translates it into reality. Term 
"neuron" in those networks has as well mathematical 
meaning. Neuron in parallel networks has its own 
technical implementation as analog and digital 
element. 
The perspectives of neuron networks utilization 
are in parallel structures for signal processing. 
Whole 2 D set of photodetectors is devided into 
several line networks.Signal analysis is carried out 
separately for each network. Simpliest way to devide 
all photodetectors into subsets is to pick out 
strings or lines at matrix. Definition of organiza- 
tion method of neurons into groups is a special 
problem which depends on concrete conditions and 
task. 
As a whole networks with nonlinear interconnec- 
tion may be used in systems for signal processing, 
to transform initial 2D optical field into different 
type of information field which should be convenient 
for further transformation and utilization. 
REFERENCES 
1. Ivanov A. Efimenko I., 1991. Multichannel 
acoustooptical device for measurements of light 
scattering in atmosphere. ICO Topical Meeting on 
Atmospheric, Volume and Surface Scattering and 
propagation. Florence : 345-347. 
2. Zakharov V., Manakov S., Novikov S., Pitaevsky L. 
1980. Soliton Theory.Inverse Scattering Method. Nauka 
Moscow. 
3. Cologero F., Degasperis A. 1985. Spectral 
Transform and Solitons. v.1, Mir, Moscow. 
Vibration of a chain with 
Phys. Soc. Japan, 22 : 431 
4, Toda M., 1967. 
nonlinear iteraction. J. 
-436. 
5. Dodd R., Eilbeck J., Gibbon J., Morris H. 1986 
Solitons and Nonlinear Wave Equations. Academic 
Press N.Y., Mir, Moscow pp. 283-287. 
6. Van de Hulst G. 1961. Light Scattering by small 
particles . Mir Moscow. 
7. Mikhaylov 6. 1987. Optimization of Monte-Karlo 
Weight Method. Nauka Moscow.