hem is 
wons at 
à those 
lculate 
w this 
supposed 
= 45 in 
ved at 
pair. 
{ferent 
nge of 
= well 
network 
ı first 
pectrum 
5 will 
5 and 
neurons 
acquire 
t pair 
tion. F 
fferent 
n some 
rameter 
ng into 
F from 
fferent 
logy of 
F will 
minimum 
ity. In 
ch. Let 
ment of 
several 
» This 
aid of 
llowing 
lines, 
s. For 
should 
tly is 
tandart 
D form 
several 
iscrete 
1, its 
2 will 
rums. If 
image 
ttering 
Fferent 
ectrum 
t light 
à small 
it into 
nations 
ponent 
|| many 
ng to 
indent. 
‘orrect 
action. 
This task may be solved with the aid of inverse 
scattering method. 
1f there are many neurons in the nonlinear 
network then coontinuous component of spectrum 
plaies more important role.It can be characterized 
by scattering koefficient RC ).But this koefficient 
can be found from the initial light distribution at 
the detector plane, because this distribution is 
Fourier transform of Rí ). Then it is necessary to 
use Helfand-Levitan equation. 
2.6 Losses in the networks 
Basic philosophy adopted in our discussion is 
absence of attenuation in network.Within some time 
period this statement is correct.But the length of 
network is proportional of l0sses. As it is 
determined in /11/, spectral components suffer 
changes with the time 
+ 
[ ydt’ 
(42) n +) -n(0)exp(- &) > r 
It means that amplitude and velocity of soliton 
reduce with cource of time.This effect may be used 
while adjusting losses at some sections of network. 
Discrete components of spectrum suffer changes due 
to attenuation too. 
3. MODEL OF NEURON 
Examine laboratory prototype of neuron.Each of 
them has receiving element.For this it is possible 
to choose capacity of photodetector.It has constant 
component CO and nonlinear one. Then all equi valent 
scheme can be represented as long line where 
inductor is an element of connection and neuron is 
nonlinear itself (see Fig.4). 
  
  
  
  
  
b) 
Figure 4, Electrical model of neuronlike 
network 
As one can see main element their is nonlinear 
capacity.This system with the aid of differential 
amplifiers produces nonlinear capacity through 
logarithmic and exponential transforms.The impedance 
of system consisting of amplifiers has capacity 
character. Inductance may be formed by convertors. 
Extent of nonlinearity was choosen in two types: 
m 
Un C = cori+« M» y 
Uo Uo 
Corresponding circuits are represented at Fig.5 a,b. 
Üne can show, that for circuit, represented at Fig.5 
C = CO In({1+ J 
C = Co (1+A) 
where À is realized as function of U 
In the case a) 
21 
ase d (t Mn) 
and in case b) 
Ra 
A= CAL) Re 
Potential UO shifts constant charges 
and enables to regulate the extent of nonlinearity. 
Electrical network consisting of such neuronlike 
elements was simulated using CAD system DISFS 
(for micro VAX type computer) to obtain real non- 
linear waves. 
When signal is small and UO is close to zero, 
network functionates as a linear one.The second type 
of nonlinearity leads to different but important 
results. There are more strong losses. Initial 
conditions were defined as potentials at all knots. 
This procedure is equivalent to carrying in non- 
linear capacities of initial charges in addition to 
constant charges.Boundary conditions were choosen as 
fixed at the ends of line. 
First type of nonlinearities produces as well 
continuous component but less than in second type. 
Simulation has shown that if nonlinearity declines 
from logarithmic form, contribution of discrete 
spectrum becomes less. If neuron is perfect and 
network consists of finite guantity of neurons then 
this network has only discrete components. 
Level of initial potential UO is important 
parameter of network.There are worked out method of 
regulating this value making use information from 
another networks. 
This idea may be realized with the aid of two 
neighbour networks.First should be network discussed 
above and second is one with elements which are 
in capacitors 
connected by linear links. This system will average 
out charge (00) at all elements. Charge ao 
determines potential UO. Furthemore let us place 
near every nonlinear network, network for production 
Qoi and connect all them together. We obtain 
ms. Qoc 
L7 1 
where n — number of neurons. 
Then whole system acquire property to estimate 
integral influence of receiving matrix. For pictures 
which produce almost homogeneous light distribution 
at the detector plane constant level of UO will 
transform optical field to high part of spectrum. If 
there are bright points at the receiving matrix, 
then they will be transformed to strong solitons. 
This method enables to pick out bright points. 
However, if there are many of those points, then 
efficiency of average charge method for all network 
decreases, In that situation it is necessary to 
devide linear networks into several groups. And 
every of them will have own average charge. It would 
enable to pick out isolated bright points and even 
lines (envisaging network as a whole). 
Third method consists in accidental connections 
of different groups of linear networks. But this 
case supposes analysis of picture using several 
sequences. Consequently it may find application only 
when change of external illuminance happens slowly. 
It should be stressed that charge DO changes 
initial conditions. 
If boundary conditions are cyclic then solitons 
move round the cycle and after some period of time 
initial picture restores. These experiments have 
shown that theoretical suggestions were close to 
characteristics of real network. 
4. EXPERIMENTAL INVESTIGATIONS 
Experimental investigations had for an object