<s>
In	O
computer	B-General_Concept
science	I-General_Concept
and	O
machine	O
learning	O
,	O
cellular	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
CNN	B-Architecture
)	O
or	O
cellular	O
nonlinear	O
networks	O
(	O
CNN	B-Architecture
)	O
are	O
a	O
parallel	B-Operating_System
computing	I-Operating_System
paradigm	O
similar	O
to	O
neural	B-General_Concept
networks	I-General_Concept
,	O
with	O
the	O
difference	O
that	O
communication	O
is	O
allowed	O
between	O
neighbouring	O
units	O
only	O
.	O
</s>
<s>
Typical	O
applications	O
include	O
image	B-Algorithm
processing	I-Algorithm
,	O
analyzing	O
3D	O
surfaces	O
,	O
solving	O
partial	O
differential	O
equations	O
,	O
reducing	O
non-visual	O
problems	O
to	O
geometric	O
maps	O
,	O
modelling	O
biological	O
vision	O
and	O
other	O
sensory-motor	O
organs	O
.	O
</s>
<s>
CNN	B-Architecture
is	O
not	O
to	O
be	O
confused	O
with	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
also	O
colloquially	O
called	O
CNN	B-Architecture
)	O
.	O
</s>
<s>
Due	O
to	O
their	O
number	O
and	O
variety	O
of	O
architectures	B-General_Concept
,	O
it	O
is	O
difficult	O
to	O
give	O
a	O
precise	O
definition	O
for	O
a	O
CNN	B-Architecture
processor	O
.	O
</s>
<s>
From	O
an	O
architecture	O
standpoint	O
,	O
CNN	B-Architecture
processors	O
are	O
a	O
system	O
of	O
finite	O
,	O
fixed-number	O
,	O
fixed-location	O
,	O
fixed-topology	O
,	O
locally	O
interconnected	O
,	O
multiple-input	O
,	O
single-output	O
,	O
nonlinear	O
processing	O
units	O
.	O
</s>
<s>
Dynamics	O
are	O
usually	O
continuous	O
,	O
as	O
in	O
the	O
case	O
of	O
Continuous-Time	O
CNN	B-Architecture
(	O
CT-CNN	O
)	O
processors	O
,	O
but	O
can	O
be	O
discrete	O
,	O
as	O
in	O
the	O
case	O
of	O
Discrete-Time	O
CNN	B-Architecture
(	O
DT-CNN	O
)	O
processors	O
.	O
</s>
<s>
a	O
Multi-Valued	O
CNN	B-Architecture
(	O
MV-CNN	O
)	O
.	O
</s>
<s>
Most	O
CNN	B-Architecture
processors	O
,	O
processing	O
units	O
are	O
identical	O
,	O
but	O
there	O
are	O
applications	O
that	O
require	O
non-identical	O
units	O
,	O
which	O
are	O
called	O
Non-Uniform	O
Processor	O
CNN	B-Architecture
(	O
NUP-CNN	O
)	O
processors	O
,	O
and	O
consist	O
of	O
different	O
types	O
of	O
cells	O
.	O
</s>
<s>
In	O
the	O
original	O
Chua-Yang	O
CNN	B-Architecture
(	O
CY-CNN	O
)	O
processor	O
,	O
the	O
state	O
of	O
the	O
cell	O
was	O
a	O
weighted	O
sum	O
of	O
the	O
inputs	O
and	O
the	O
output	O
was	O
a	O
piecewise	O
linear	O
function	O
.	O
</s>
<s>
However	O
,	O
like	O
the	O
original	O
perceptron-based	O
neural	B-General_Concept
networks	I-General_Concept
,	O
the	O
functions	O
it	O
could	O
perform	O
were	O
limited	O
:	O
specifically	O
,	O
it	O
was	O
incapable	O
of	O
modeling	O
non-linear	O
functions	O
,	O
such	O
as	O
XOR	B-Application
.	O
</s>
<s>
More	O
complex	O
functions	O
are	O
realizable	O
via	O
Non-Linear	O
CNN	B-Architecture
(	O
NL-CNN	O
)	O
processors	O
.	O
</s>
<s>
Topologically	B-Architecture
,	O
cells	O
can	O
be	O
arranged	O
on	O
an	O
infinite	O
plane	O
or	O
on	O
a	O
toroidal	O
space	O
.	O
</s>
<s>
Cell	O
interconnect	O
is	O
local	O
,	O
meaning	O
that	O
all	O
connections	O
between	O
cells	O
are	O
within	O
a	O
specified	O
radius	O
(	O
with	O
distance	O
measured	O
topologically	B-Architecture
)	O
.	O
</s>
<s>
Most	O
CNN	B-Architecture
architectures	B-General_Concept
have	O
cells	O
with	O
the	O
same	O
relative	O
interconnects	O
,	O
but	O
there	O
are	O
applications	O
that	O
require	O
a	O
spatially	O
variant	O
topology	B-Architecture
,	O
i.e.	O
</s>
<s>
Multiple-Neighborhood-Size	O
CNN	B-Architecture
(	O
MNS-CNN	O
)	O
processors	O
.	O
</s>
<s>
Also	O
,	O
Multiple-Layer	O
CNN	B-Architecture
(	O
ML-CNN	O
)	O
processors	O
,	O
where	O
all	O
cells	O
on	O
the	O
same	O
layer	O
are	O
identical	O
,	O
can	O
be	O
used	O
to	O
extend	O
the	O
capability	O
of	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
In	O
this	O
sense	O
,	O
CNN	B-Architecture
processors	O
are	O
systems	O
because	O
their	O
dynamics	O
are	O
derived	O
from	O
the	O
interaction	O
between	O
the	O
processing	O
units	O
and	O
not	O
within	O
processing	O
units	O
.	O
</s>
<s>
Mathematically	O
,	O
the	O
relationship	O
between	O
a	O
cell	O
and	O
its	O
neighbors	O
,	O
located	O
within	O
an	O
area	O
of	O
influence	O
,	O
can	O
be	O
defined	O
by	O
a	O
coupling	B-Application
law	O
,	O
and	O
this	O
is	O
what	O
primarily	O
determines	O
the	O
behavior	O
of	O
the	O
processor	O
.	O
</s>
<s>
When	O
the	O
coupling	B-Application
laws	O
are	O
modeled	O
by	O
fuzzy	O
logic	O
,	O
it	O
is	O
a	O
fuzzy	O
CNN	B-Architecture
.	O
</s>
<s>
When	O
these	O
laws	O
are	O
modeled	O
by	O
computational	O
verb	O
logic	O
,	O
it	O
becomes	O
a	O
computational	O
verb	O
CNN	B-Architecture
.	O
</s>
<s>
Both	O
fuzzy	O
and	O
verb	O
CNNs	B-Architecture
are	O
useful	O
for	O
modelling	O
social	O
networks	O
when	O
the	O
local	O
couplings	B-Application
are	O
achieved	O
by	O
linguistic	O
terms	O
.	O
</s>
<s>
The	O
idea	O
of	O
CNN	B-Architecture
processors	O
was	O
introduced	O
by	O
Leon	O
Chua	O
and	O
Lin	O
Yang	O
in	O
1988	O
.	O
</s>
<s>
In	O
these	O
articles	O
,	O
Chua	O
and	O
Yang	O
outline	O
the	O
underlying	O
mathematics	O
behind	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
They	O
use	O
this	O
mathematical	O
model	O
to	O
demonstrate	O
,	O
for	O
a	O
specific	O
CNN	B-Architecture
implementation	O
,	O
that	O
if	O
the	O
inputs	O
are	O
static	O
,	O
the	O
processing	O
units	O
will	O
converge	O
,	O
and	O
can	O
be	O
used	O
to	O
perform	O
useful	O
calculations	O
.	O
</s>
<s>
They	O
then	O
suggest	O
one	O
of	O
the	O
first	O
applications	O
of	O
CNN	B-Architecture
processors	O
:	O
image	B-Algorithm
processing	I-Algorithm
and	O
pattern	O
recognition	O
(	O
which	O
is	O
still	O
the	O
largest	O
application	O
to	O
date	O
)	O
.	O
</s>
<s>
Leon	O
Chua	O
is	O
still	O
active	O
in	O
CNN	B-Architecture
research	O
and	O
publishes	O
many	O
of	O
his	O
articles	O
in	O
the	O
International	O
Journal	O
of	O
Bifurcation	O
and	O
Chaos	O
,	O
of	O
which	O
he	O
is	O
an	O
editor	O
.	O
</s>
<s>
Both	O
IEEE	O
Transactions	O
on	O
Circuits	O
and	O
Systems	O
and	O
the	O
International	O
Journal	O
of	O
Bifurcation	O
also	O
contain	O
a	O
variety	O
of	O
useful	O
articles	O
on	O
CNN	B-Architecture
processors	O
authored	O
by	O
other	O
knowledgeable	O
researchers	O
.	O
</s>
<s>
The	O
former	O
tends	O
to	O
focus	O
on	O
new	O
CNN	B-Architecture
architectures	B-General_Concept
and	O
the	O
latter	O
more	O
on	O
the	O
dynamical	O
aspects	O
of	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
In	O
1993	O
,	O
Tamas	O
Roska	O
and	O
Leon	O
Chua	O
introduced	O
the	O
first	O
algorithmically	O
programmable	O
analog	O
CNN	B-Architecture
processor	O
in	O
the	O
world	O
.	O
</s>
<s>
This	O
article	O
proved	O
that	O
CNN	B-Architecture
processors	O
were	O
producible	O
and	O
provided	O
researchers	O
a	O
physical	O
platform	O
to	O
test	O
their	O
CNN	B-Architecture
theories	O
.	O
</s>
<s>
After	O
this	O
article	O
,	O
companies	O
started	O
to	O
invest	O
into	O
larger	O
,	O
more	O
capable	O
processors	O
,	O
based	O
on	O
the	O
same	O
basic	O
architecture	O
as	O
the	O
CNN	B-Architecture
Universal	O
Processor	O
.	O
</s>
<s>
Tamas	O
Roska	O
is	O
another	O
key	O
contributor	O
to	O
CNNs	B-Architecture
.	O
</s>
<s>
His	O
name	O
is	O
often	O
associated	O
with	O
biologically	O
inspired	O
information	O
processing	O
platforms	O
and	O
algorithms	O
,	O
and	O
he	O
has	O
published	O
numerous	O
key	O
articles	O
and	O
has	O
been	O
involved	O
with	O
companies	O
and	O
research	O
institutions	O
developing	O
CNN	B-Architecture
technology	O
.	O
</s>
<s>
Two	O
references	O
are	O
considered	O
invaluable	O
since	O
they	O
manage	O
to	O
organize	O
the	O
vast	O
amount	O
of	O
CNN	B-Architecture
literature	O
into	O
a	O
coherent	O
framework	O
:	O
</s>
<s>
The	O
paper	O
provides	O
a	O
concise	O
intro	O
to	O
definitions	O
,	O
CNN	B-Architecture
types	O
,	O
dynamics	O
,	O
implementations	O
,	O
and	O
applications	O
.	O
</s>
<s>
"	O
Cellular	B-Algorithm
Neural	I-Algorithm
Networks	I-Algorithm
and	O
Visual	O
Computing	O
Foundations	O
and	O
Applications	O
"	O
,	O
written	O
by	O
Leon	O
Chua	O
and	O
Tamas	O
Roska	O
,	O
which	O
provides	O
examples	O
and	O
exercises	O
.	O
</s>
<s>
The	O
book	O
covers	O
many	O
different	O
aspects	O
of	O
CNN	B-Architecture
processors	O
and	O
can	O
serve	O
as	O
a	O
textbook	O
for	O
a	O
Masters	O
or	O
Ph.D.	O
course	O
.	O
</s>
<s>
The	O
proceedings	O
of	O
"	O
The	O
International	O
Workshop	O
on	O
Cellular	B-Algorithm
Neural	I-Algorithm
Networks	I-Algorithm
and	O
Their	O
Applications	O
"	O
provide	O
much	O
CNN	B-Architecture
literature	O
.	O
</s>
<s>
For	O
an	O
understanding	O
of	O
the	O
analog	O
semiconductor	O
based	O
CNN	B-Architecture
technology	O
,	O
AnaLogic	O
Computers	O
has	O
their	O
product	O
line	O
,	O
in	O
addition	O
to	O
the	O
published	O
articles	O
available	O
on	O
their	O
homepage	O
and	O
their	O
publication	O
list	O
.	O
</s>
<s>
They	O
also	O
have	O
information	O
on	O
other	O
CNN	B-Architecture
technologies	O
such	O
as	O
optical	O
computing	O
.	O
</s>
<s>
Many	O
of	O
the	O
commonly	O
used	O
functions	O
have	O
already	O
been	O
implemented	O
using	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
A	O
good	O
reference	O
point	O
for	O
some	O
of	O
these	O
can	O
be	O
found	O
in	O
image	B-Algorithm
processing	I-Algorithm
libraries	O
for	O
CNN	B-Architecture
based	O
visual	O
computers	O
such	O
as	O
Analogic	O
’s	O
CNN-based	O
systems	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
could	O
be	O
thought	O
of	O
as	O
a	O
hybrid	O
between	O
ANN	B-Architecture
and	O
Continuous	O
Automata	O
(	O
CA	O
)	O
.	O
</s>
<s>
The	O
processing	O
units	O
of	O
CNN	B-Architecture
and	O
NN	O
are	O
similar	O
.	O
</s>
<s>
However	O
,	O
in	O
CNN	B-Architecture
processors	O
,	O
connections	O
are	O
made	O
locally	O
,	O
whereas	O
in	O
ANN	B-Architecture
,	O
connections	O
are	O
global	O
.	O
</s>
<s>
For	O
example	O
,	O
neurons	O
in	O
one	O
layer	O
are	O
fully	O
connected	O
to	O
another	O
layer	O
in	O
a	O
feed-forward	B-Algorithm
NN	I-Algorithm
and	O
all	O
the	O
neurons	O
are	O
fully	O
interconnected	O
in	O
Hopfield	B-Algorithm
networks	I-Algorithm
.	O
</s>
<s>
But	O
in	O
CNN	B-Architecture
processors	O
,	O
the	O
weights	O
are	O
used	O
to	O
determine	O
the	O
dynamics	O
of	O
the	O
system	O
.	O
</s>
<s>
Furthermore	O
,	O
due	O
to	O
the	O
high	O
inter-connectivity	O
of	O
ANNs	O
,	O
they	O
tend	O
not	O
exploit	O
locality	O
in	O
either	O
the	O
data	O
set	O
or	O
the	O
processing	O
and	O
as	O
a	O
result	O
,	O
they	O
usually	O
are	O
highly	O
redundant	O
systems	O
that	O
allow	O
for	O
robust	B-Application
,	O
fault-tolerant	O
behavior	O
without	O
catastrophic	O
errors	O
.	O
</s>
<s>
A	O
cross	O
between	O
an	O
ANN	B-Architecture
and	O
a	O
CNN	B-Architecture
processor	O
is	O
a	O
Ratio	O
Memory	O
CNN	B-Architecture
(	O
RMCNN	O
)	O
.	O
</s>
<s>
In	O
RMCNN	O
processors	O
,	O
the	O
cell	O
interconnect	O
is	O
local	O
and	O
topologically	B-Architecture
invariant	O
,	O
but	O
the	O
weights	O
are	O
used	O
to	O
store	O
previous	O
states	O
and	O
not	O
to	O
control	O
dynamics	O
.	O
</s>
<s>
The	O
topology	B-Architecture
and	O
dynamics	O
of	O
CNN	B-Architecture
processors	O
closely	O
resembles	O
that	O
of	O
CA	O
.	O
</s>
<s>
Like	O
most	O
CNN	B-Architecture
processors	O
,	O
CA	O
consists	O
of	O
a	O
fixed-number	O
of	O
identical	O
processors	O
that	O
are	O
spatially	O
discrete	O
and	O
topologically	B-Architecture
uniform	O
.	O
</s>
<s>
The	O
difference	O
is	O
that	O
most	O
CNN	B-Architecture
processors	O
are	O
continuous-valued	O
whereas	O
CA	O
have	O
discrete-values	O
.	O
</s>
<s>
Furthermore	O
,	O
the	O
CNN	B-Architecture
processor	O
's	O
cell	O
behavior	O
is	O
defined	O
via	O
some	O
non-linear	O
function	O
whereas	O
CA	O
processor	O
cells	O
are	O
defined	O
by	O
some	O
state	O
machine	O
.	O
</s>
<s>
Depending	O
on	O
how	O
a	O
given	O
Continuous	O
Automata	O
is	O
specified	O
,	O
it	O
can	O
also	O
be	O
a	O
CNN	B-Architecture
.	O
</s>
<s>
Continuous	O
Spatial	O
Automata	O
processors	O
can	O
be	O
physically	O
realized	O
though	O
an	O
unconventional	O
information	O
processing	O
platform	O
such	O
as	O
a	O
chemical	B-Device
computer	I-Device
.	O
</s>
<s>
Furthermore	O
,	O
it	O
is	O
conceivable	O
that	O
large	O
CNN	B-Architecture
processors	O
(	O
in	O
terms	O
of	O
the	O
resolution	O
of	O
the	O
input	O
and	O
output	O
)	O
can	O
be	O
modeled	O
as	O
a	O
Continuous	O
Spatial	O
Automata	O
.	O
</s>
<s>
The	O
dynamical	O
behavior	O
of	O
CNN	B-Architecture
processors	O
can	O
be	O
expressed	O
using	O
differential	O
equations	O
,	O
where	O
each	O
equation	O
represents	O
the	O
state	O
of	O
an	O
individual	O
processing	O
unit	O
.	O
</s>
<s>
The	O
behavior	O
of	O
the	O
entire	O
CNN	B-Architecture
processor	O
is	O
defined	O
by	O
its	O
initial	O
conditions	O
,	O
inputs	O
,	O
cell	O
interconnections	O
(	O
topology	B-Architecture
and	O
weights	O
)	O
,	O
and	O
the	O
cells	O
themselves	O
.	O
</s>
<s>
One	O
possible	O
use	O
of	O
CNN	B-Architecture
processors	O
is	O
to	O
generate	O
and	O
respond	O
to	O
signals	O
of	O
specific	O
dynamical	O
properties	O
.	O
</s>
<s>
For	O
example	O
,	O
CNN	B-Architecture
processors	O
have	O
been	O
used	O
to	O
generate	O
multiscroll	O
chaos	O
(	O
like	O
the	O
Chen	O
attractor	O
)	O
,	O
synchronize	O
with	O
chaotic	O
systems	O
,	O
and	O
exhibit	O
multi-level	O
hysteresis	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
are	O
designed	O
specifically	O
to	O
solve	O
local	O
,	O
low-level	O
,	O
processor	O
intensive	O
problems	O
expressed	O
as	O
a	O
function	O
of	O
space	O
and	O
time	O
.	O
</s>
<s>
For	O
example	O
,	O
CNN	B-Architecture
processors	O
can	O
be	O
used	O
to	O
implement	O
high-pass	O
and	O
low-pass	O
filters	O
and	O
morphological	B-Algorithm
operators	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
can	O
be	O
used	O
as	O
Reaction-Diffusion	O
(	O
RD	O
)	O
processors	O
.	O
</s>
<s>
RD	O
processors	O
are	O
spatially	O
invariant	O
,	O
topologically	B-Architecture
invariant	O
,	O
analog	O
,	O
parallel	B-Operating_System
processors	I-Operating_System
characterized	O
by	O
reactions	O
,	O
where	O
two	O
agents	O
can	O
combine	O
to	O
create	O
a	O
third	O
agent	O
,	O
and	O
diffusion	O
,	O
the	O
spreading	O
of	O
agents	O
.	O
</s>
<s>
RD	O
processors	O
are	O
typically	O
implemented	O
through	O
chemicals	O
in	O
a	O
Petri	O
dish	O
(	O
processor	O
)	O
,	O
light	O
(	O
input	O
)	O
,	O
and	O
a	O
camera	O
(	O
output	O
)	O
however	O
RD	O
processors	O
can	O
also	O
be	O
implemented	O
through	O
a	O
multi-layer	O
CNN	B-Architecture
processor	O
.	O
</s>
<s>
D	O
processors	O
can	O
be	O
used	O
to	O
create	O
Voronoi	B-Architecture
diagrams	I-Architecture
and	O
perform	O
skeletonisation	O
.	O
</s>
<s>
The	O
main	O
difference	O
between	O
the	O
chemical	O
implementation	O
and	O
the	O
CNN	B-Architecture
implementation	O
is	O
that	O
CNN	B-Architecture
implementations	O
are	O
considerably	O
faster	O
than	O
their	O
chemical	O
counterparts	O
and	O
chemical	O
processors	O
are	O
spatially	O
continuous	O
whereas	O
the	O
CNN	B-Architecture
processors	O
are	O
spatially	O
discrete	O
.	O
</s>
<s>
The	O
most	O
researched	O
RD	O
processor	O
,	O
Belousov-Zhabotinsky	O
(	O
BZ	O
)	O
processors	O
,	O
has	O
already	O
been	O
simulated	O
using	O
a	O
four-layer	O
CNN	B-Architecture
processors	O
and	O
has	O
been	O
implemented	O
in	O
a	O
semiconductor	O
.	O
</s>
<s>
The	O
applications	O
of	O
CNNs	B-Architecture
to	O
Boolean	O
functions	O
is	O
discussed	O
in	O
the	O
paper	O
by	O
Fangyue	O
Chen	O
,	O
Guolong	O
He	O
,	O
Xiubin	O
Xu	O
,	O
and	O
Guanrong	O
Chen	O
,	O
"	O
Implementation	O
of	O
Arbitrary	O
Boolean	O
Functions	O
via	O
CNN	B-Architecture
"	O
.	O
</s>
<s>
Although	O
CNN	B-Architecture
processors	O
are	O
primarily	O
intended	O
for	O
analog	O
calculations	O
,	O
certain	O
types	O
of	O
CNN	B-Architecture
processors	O
can	O
implement	O
any	O
Boolean	O
function	O
,	O
allowing	O
simulating	O
CA	O
.	O
</s>
<s>
Since	O
some	O
CA	O
are	O
Universal	O
Turing	B-Architecture
machines	I-Architecture
(	O
UTM	O
)	O
,	O
capable	O
of	O
simulating	O
any	O
algorithm	O
can	O
be	O
performed	O
on	O
processors	O
based	O
on	O
the	O
von	B-Architecture
Neumann	I-Architecture
architecture	I-Architecture
,	O
that	O
makes	O
this	O
type	O
of	O
CNN	B-Architecture
processors	O
,	O
universal	O
CNN	B-Architecture
,	O
a	O
UTM	O
.	O
</s>
<s>
One	O
CNN	B-Architecture
architecture	O
consists	O
of	O
an	O
additional	O
layer	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
resulted	O
in	O
the	O
simplest	O
realization	O
of	O
Conway	O
’s	O
Game	O
of	O
Life	O
and	O
Wolfram	O
’s	O
Rule	O
110	O
,	O
the	O
simplest	O
known	O
universal	O
Turing	B-Architecture
Machine	I-Architecture
.	O
</s>
<s>
This	O
unique	O
,	O
dynamical	O
representation	O
of	O
an	O
old	O
systems	O
,	O
allows	O
researchers	O
to	O
apply	O
techniques	O
and	O
hardware	O
developed	O
for	O
CNN	B-Architecture
to	O
better	O
understand	O
important	O
CA	O
.	O
</s>
<s>
Furthermore	O
,	O
the	O
continuous	O
state	O
space	O
of	O
CNN	B-Architecture
processors	O
,	O
with	O
slight	O
modifications	O
that	O
have	O
no	O
equivalent	O
in	O
Cellular	O
Automata	O
,	O
creates	O
emergent	O
behavior	O
never	O
seen	O
before	O
.	O
</s>
<s>
Any	O
information	O
processing	O
platform	O
that	O
allows	O
the	O
construction	O
of	O
arbitrary	O
Boolean	O
functions	O
is	O
called	O
universal	O
,	O
and	O
as	O
result	O
,	O
this	O
class	O
CNN	B-Architecture
processors	O
are	O
commonly	O
referred	O
to	O
as	O
universal	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
The	O
original	O
CNN	B-Architecture
processors	O
can	O
only	O
perform	O
linearly	O
separable	O
Boolean	O
functions	O
.	O
</s>
<s>
By	O
translating	O
functions	O
from	O
digital	O
logic	O
or	O
look-up	O
table	O
domains	O
into	O
the	O
CNN	B-Architecture
domain	O
,	O
some	O
functions	O
can	O
be	O
considerably	O
simplified	O
.	O
</s>
<s>
Not	O
only	O
is	O
there	O
a	O
reduction	O
in	O
the	O
function	O
complexity	O
,	O
but	O
the	O
CNN	B-Architecture
implementation	O
parameters	O
can	O
be	O
represented	O
in	O
the	O
continuous	O
,	O
real-number	O
domain	O
.	O
</s>
<s>
There	O
are	O
two	O
methods	O
by	O
which	O
to	O
select	O
a	O
CNN	B-Architecture
processor	O
along	O
with	O
a	O
template	O
or	O
weights	O
.	O
</s>
<s>
Researchers	O
have	O
used	O
back-propagation	B-Algorithm
and	O
genetic	B-Algorithm
algorithms	I-Algorithm
to	O
learn	O
and	O
perform	O
functions	O
.	O
</s>
<s>
Back-propagation	B-Algorithm
algorithms	O
tend	O
to	O
be	O
faster	O
,	O
but	O
genetic	B-Algorithm
algorithms	I-Algorithm
are	O
useful	O
because	O
they	O
provide	O
a	O
mechanism	O
to	O
find	O
a	O
solution	O
in	O
a	O
discontinuous	O
,	O
noisy	O
search	O
space	O
.	O
</s>
<s>
There	O
are	O
toy	O
models	O
simulating	O
CNN	B-Architecture
processors	O
using	O
billiard	O
balls	O
,	O
but	O
these	O
are	O
used	O
for	O
theoretical	O
studies	O
.	O
</s>
<s>
In	O
practice	O
,	O
CNN	B-Architecture
are	O
physically	O
implemented	O
on	O
hardware	O
and	O
current	O
technologies	O
such	O
as	O
semiconductors	O
.	O
</s>
<s>
There	O
are	O
plans	O
to	O
migrate	O
CNN	B-Architecture
processors	O
to	O
emerging	O
technologies	O
in	O
the	O
future	O
.	O
</s>
<s>
Semiconductor-based	O
CNN	B-Architecture
processors	O
can	O
be	O
segmented	O
into	O
analog	O
CNN	B-Architecture
processors	O
,	O
digital	O
CNN	B-Architecture
processors	O
,	O
and	O
CNN	B-Architecture
processors	O
emulated	B-Application
using	O
digital	O
processors	O
.	O
</s>
<s>
Analog	O
CNN	B-Architecture
processors	O
were	O
the	O
first	O
to	O
be	O
developed	O
.	O
</s>
<s>
Analog	O
CNN	B-Architecture
processors	O
share	O
some	O
of	O
the	O
same	O
advantages	O
as	O
their	O
predecessors	O
,	O
specifically	O
speed	O
.	O
</s>
<s>
The	O
first	O
analog	O
CNN	B-Architecture
processors	O
were	O
able	O
to	O
perform	O
real-time	O
ultra-high	O
frame-rate	O
(	O
>10	O
,	O
000	O
frame/s	O
)	O
processing	O
unachievable	O
by	O
digital	O
processors	O
.	O
</s>
<s>
The	O
analog	O
implementation	O
of	O
CNN	B-Architecture
processors	O
requires	O
less	O
area	O
and	O
consumes	O
less	O
power	O
than	O
their	O
digital	O
counterparts	O
.	O
</s>
<s>
Although	O
the	O
accuracy	O
of	O
analog	O
CNN	B-Architecture
processors	O
does	O
not	O
compare	O
to	O
their	O
digital	O
counterparts	O
,	O
for	O
many	O
applications	O
,	O
noise	O
and	O
process	O
variances	O
are	O
small	O
enough	O
not	O
to	O
perceptually	O
affect	O
the	O
image	O
quality	O
.	O
</s>
<s>
The	O
first	O
algorithmically	O
programmable	O
,	O
analog	O
CNN	B-Architecture
processor	O
was	O
created	O
in	O
1993	O
.	O
</s>
<s>
It	O
was	O
named	O
the	O
CNN	B-Architecture
Universal	O
Processor	O
because	O
its	O
internal	O
controller	O
allowed	O
multiple	O
templates	O
to	O
be	O
performed	O
on	O
the	O
same	O
data	O
set	O
,	O
thus	O
simulating	O
multiple	O
layers	O
and	O
allowing	O
for	O
universal	B-Architecture
computation	I-Architecture
.	O
</s>
<s>
The	O
processor	O
was	O
developed	O
in	O
order	O
to	O
determine	O
CNN	B-Architecture
processor	O
producibility	O
and	O
utility	O
.	O
</s>
<s>
The	O
CNN	B-Architecture
concept	O
proved	O
promising	O
and	O
by	O
2000	O
,	O
there	O
were	O
at	O
least	O
six	O
organizations	O
designing	O
algorithmically	O
programmable	O
,	O
analog	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
In	O
the	O
2000s	O
,	O
AnaFocus	O
,	O
a	O
mixed-signal	O
semiconductor	O
company	O
from	O
the	O
University	O
of	O
Seville	O
,	O
introduced	O
their	O
ACE	O
prototype	O
CNN	B-Architecture
processor	O
product	O
line	O
.	O
</s>
<s>
AnaFocus	O
also	O
had	O
a	O
multilayer	O
CASE	O
prototype	O
CNN	B-Architecture
processors	O
line	O
.	O
</s>
<s>
Another	O
company	O
,	O
AnaLogic	O
Computers	O
was	O
founded	O
in	O
2000	O
by	O
many	O
of	O
the	O
same	O
researchers	O
behind	O
the	O
first	O
algorithmically	O
programmable	O
CNN	B-Architecture
Universal	O
Processor	O
.	O
</s>
<s>
This	O
allowed	O
CNN	B-Architecture
processing	O
to	O
be	O
easily	O
included	O
in	O
a	O
desktop	O
computer	O
.	O
</s>
<s>
Some	O
researchers	O
developed	O
their	O
own	O
custom	O
analog	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
Chung-Yu	O
Wu	O
and	O
Chiu-Hung	O
Cheng	O
from	O
National	O
Chiao	O
Tung	O
University	O
designed	O
a	O
RM-CNN	O
processor	O
to	O
learn	O
more	O
about	O
pattern	O
learning	O
and	O
recognition	O
.	O
</s>
<s>
Researchers	O
from	O
the	O
National	O
Lien-Ho	O
Institute	O
of	O
Technology	O
(	O
W	O
.	O
Yen	O
,	O
R	O
.	O
Chen	O
and	O
J	O
.	O
Lai	O
)	O
developed	O
a	O
Min-Max	O
CNN	B-Architecture
(	O
MMCNN	O
)	O
processor	O
to	O
learn	O
more	O
about	O
CNN	B-Architecture
dynamics	O
.	O
</s>
<s>
Despite	O
their	O
speed	O
and	O
low	O
power	O
consumption	O
,	O
there	O
are	O
some	O
significant	O
drawbacks	O
to	O
analog	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
First	O
,	O
analog	O
CNN	B-Architecture
processors	O
can	O
potentially	O
create	O
erroneous	O
results	O
due	O
to	O
environment	O
and	O
process	O
variation	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
chaotic	O
communication	O
,	O
process	O
variation	O
will	O
change	O
the	O
trajectory	B-Application
of	O
a	O
given	O
system	O
in	O
phase	O
space	O
,	O
resulting	O
in	O
a	O
loss	O
of	O
synchronicity/stability	O
.	O
</s>
<s>
Other	O
researchers	O
are	O
improving	O
the	O
semiconductor	O
process	O
to	O
more	O
closely	O
match	O
theoretical	O
CNN	B-Architecture
performance	O
.	O
</s>
<s>
Other	O
researchers	O
are	O
investigating	O
different	O
,	O
potentially	O
more	O
robust	B-Application
CNN	B-Architecture
architectures	B-General_Concept
.	O
</s>
<s>
Not	O
only	O
does	O
process	O
variation	O
limit	O
what	O
can	O
be	O
done	O
with	O
current	O
analog	O
CNN	B-Architecture
processors	O
,	O
it	O
is	O
also	O
a	O
barrier	O
for	O
creating	O
more	O
complex	O
processing	O
units	O
.	O
</s>
<s>
cannot	O
be	O
implemented	O
in	O
a	O
real-time	O
analog	O
CNN	B-Architecture
processor	O
.	O
</s>
<s>
Also	O
,	O
the	O
semiconductor	O
"	O
real	O
estate	O
"	O
for	O
processing	O
units	O
limits	O
the	O
size	O
of	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
Currently	O
the	O
largest	O
AnaVision	O
CNN-based	O
vision	O
processor	O
consists	O
of	O
a	O
4K	O
detector	O
,	O
which	O
is	O
significantly	O
less	O
than	O
the	O
megapixel	O
detectors	O
found	O
in	O
affordable	O
,	O
consumer	O
cameras	O
.	O
</s>
<s>
Also	O
,	O
CNN	B-Architecture
processor	O
architecture	O
is	O
being	O
re-evaluated	O
.	O
</s>
<s>
For	O
example	O
,	O
Star-CNN	O
processors	O
,	O
where	O
one	O
analog	O
multiplier	O
is	O
time-shared	O
between	O
multiple	O
processor	O
units	O
,	O
have	O
been	O
proposed	O
and	O
are	O
expected	O
to	O
result	O
in	O
processor	O
unit	O
reduction	O
size	O
of	O
80%	O
.	O
</s>
<s>
Although	O
not	O
nearly	O
as	O
fast	O
and	O
energy	O
efficient	O
,	O
digital	O
CNN	B-Architecture
processors	O
do	O
not	O
share	O
the	O
problems	O
of	O
process	O
variation	O
and	O
feature	O
size	O
of	O
their	O
analog	O
counterparts	O
.	O
</s>
<s>
This	O
allows	O
digital	O
CNN	B-Architecture
processors	O
to	O
include	O
nested	O
processor	O
units	O
,	O
non-linearities	O
,	O
etc	O
.	O
</s>
<s>
In	O
addition	O
,	O
digital	O
CNN	B-Architecture
are	O
more	O
flexible	O
,	O
cost	O
less	O
and	O
are	O
easier	O
to	O
integrate	O
.	O
</s>
<s>
The	O
most	O
common	O
implementation	O
of	O
digital	O
CNN	B-Architecture
processors	O
uses	O
an	O
FPGA	B-Architecture
.	O
</s>
<s>
Eutecus	O
,	O
founded	O
in	O
2002	O
and	O
operating	O
in	O
Berkeley	O
,	O
provides	O
intellectual	O
property	O
that	O
can	O
be	O
synthesized	O
into	O
an	O
Altera	O
FPGA	B-Architecture
.	O
</s>
<s>
Their	O
digital	O
320x280	O
,	O
FPGA-based	O
CNN	B-Architecture
processors	O
run	O
at	O
30	O
frame/s	O
and	O
there	O
are	O
plans	O
to	O
make	O
a	O
fast	O
digital	O
ASIC	O
.	O
</s>
<s>
Eustecus	O
is	O
a	O
strategic	O
partner	O
of	O
AnaLogic	O
computers	O
,	O
and	O
their	O
FPGA	B-Architecture
designs	O
can	O
be	O
found	O
in	O
several	O
of	O
AnaLogic	O
’s	O
products	O
.	O
</s>
<s>
Eutecus	O
is	O
also	O
developing	O
software	O
libraries	O
to	O
perform	O
tasks	O
including	O
but	O
not	O
limited	O
to	O
video	O
analytics	O
for	O
the	O
video	O
security	O
market	O
,	O
feature	O
classification	O
,	O
multi-target	O
tracking	O
,	O
signal	O
and	O
image	B-Algorithm
processing	I-Algorithm
and	O
flow	O
processing	O
.	O
</s>
<s>
Many	O
of	O
these	O
routines	O
are	O
derived	O
using	O
CNN-like	O
processing	O
.	O
</s>
<s>
For	O
those	O
wanting	O
to	O
perform	O
CNN	B-Architecture
simulations	O
for	O
prototyping	O
,	O
low-speed	O
applications	O
,	O
or	O
research	O
,	O
there	O
are	O
several	O
options	O
.	O
</s>
<s>
First	O
,	O
there	O
are	O
precise	O
CNN	B-Architecture
emulation	B-Application
software	I-Application
packages	O
like	O
SCNN	O
2000	O
.	O
</s>
<s>
If	O
the	O
speed	O
is	O
prohibitive	O
,	O
there	O
are	O
mathematical	O
techniques	O
,	O
such	O
as	O
Jacobi	O
’s	O
Iterative	O
Method	O
or	O
Forward-Backward	O
Recursions	O
that	O
can	O
be	O
used	O
to	O
derive	O
the	O
steady	O
state	O
solution	O
of	O
a	O
CNN	B-Architecture
processor	O
.	O
</s>
<s>
Lastly	O
,	O
digital	O
CNN	B-Architecture
processors	O
can	O
be	O
emulated	B-Application
on	O
highly	O
parallel	O
,	O
application-specific	O
processors	O
,	O
such	O
as	O
graphics	O
processors	O
.	O
</s>
<s>
Implementing	O
neural	B-General_Concept
networks	I-General_Concept
using	O
graphics	O
processors	O
is	O
an	O
area	O
of	O
further	O
research	O
.	O
</s>
<s>
Researchers	O
are	O
also	O
perusing	O
alternate	O
technologies	O
for	O
CNN	B-Architecture
processors	O
.	O
</s>
<s>
Although	O
current	O
CNN	B-Architecture
processors	O
circumvent	O
some	O
of	O
the	O
problems	O
associated	O
with	O
their	O
digital	O
counterparts	O
,	O
they	O
do	O
share	O
some	O
of	O
the	O
same	O
long-term	O
problems	O
common	O
to	O
all	O
semiconductor-based	O
processors	O
.	O
</s>
<s>
AnaLogic	O
Computers	O
,	O
is	O
developing	O
optical	O
CNN	B-Architecture
processors	O
,	O
which	O
combine	O
optics	O
,	O
lasers	O
,	O
and	O
biological	O
and	O
holographic	B-Algorithm
memories	O
.	O
</s>
<s>
What	O
initially	O
was	O
technology	O
exploration	O
resulted	O
in	O
a	O
500x500	O
CNN	B-Architecture
processor	O
able	O
to	O
perform	O
300	O
giga-operations	O
per	O
second	O
.	O
</s>
<s>
Another	O
promising	O
technology	O
for	O
CNN	B-Architecture
processors	O
is	O
nanotechnology	O
.	O
</s>
<s>
One	O
nanotechnology	O
concept	O
being	O
investigated	O
is	O
using	O
single	O
electron	O
tunneling	O
junctions	O
,	O
which	O
can	O
be	O
made	O
into	O
single-electron	O
or	O
high-current	O
transistors	O
,	O
to	O
create	O
McCulloch-Pitts	O
CNN	B-Architecture
processing	O
units	O
.	O
</s>
<s>
In	O
summary	O
,	O
CNN	B-Architecture
processors	O
have	O
been	O
implemented	O
and	O
provide	O
value	O
to	O
their	O
users	O
.	O
</s>
<s>
Researchers	O
are	O
also	O
transitioning	O
CNN	B-Architecture
processors	O
into	O
emerging	O
technologies	O
.	O
</s>
<s>
Therefore	O
,	O
if	O
the	O
CNN	B-Architecture
architecture	O
is	O
suited	O
for	O
a	O
specific	O
information	O
processing	O
system	O
,	O
there	O
are	O
processors	O
available	O
for	O
purchase	O
(	O
as	O
there	O
will	O
be	O
for	O
the	O
foreseeable	O
future	O
)	O
.	O
</s>
<s>
CNN	B-Architecture
researchers	O
have	O
diverse	O
interests	O
,	O
ranging	O
from	O
physical	O
,	O
engineering	O
,	O
theoretical	O
,	O
mathematical	O
,	O
computational	O
,	O
and	O
philosophical	O
applications	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
were	O
designed	O
to	O
perform	O
image	B-Algorithm
processing	I-Algorithm
;	O
specifically	O
,	O
real-time	O
ultra-high	O
frame-rate	O
(	O
>10	O
,	O
000	O
frame/s	O
)	O
processing	O
for	O
applications	O
like	O
particle	O
detection	O
in	O
jet	O
engine	O
fluids	O
and	O
spark-plug	O
detection	O
.	O
</s>
<s>
Currently	O
,	O
CNN	B-Architecture
processors	O
can	O
achieve	O
up	O
to	O
50,000	O
frames	O
per	O
second	O
,	O
and	O
for	O
certain	O
applications	O
such	O
as	O
missile	O
tracking	O
,	O
flash	O
detection	O
,	O
and	O
spark-plug	O
diagnostics	O
these	O
microprocessors	O
have	O
outperformed	O
a	O
conventional	O
supercomputer	B-Architecture
.	O
</s>
<s>
CNN	B-Architecture
processors	O
lend	O
themselves	O
to	O
local	O
,	O
low-level	O
,	O
processor	O
intensive	O
operations	O
and	O
have	O
been	O
used	O
in	O
feature	O
extraction	O
,	O
level	O
and	O
gain	O
adjustments	O
,	O
color	O
constancy	O
detection	O
,	O
contrast	O
enhancement	O
,	O
deconvolution	B-Algorithm
,	O
image	B-General_Concept
compression	I-General_Concept
,	O
motion	O
estimation	O
,	O
image	O
encoding	O
,	O
image	O
decoding	O
,	O
image	B-Algorithm
segmentation	I-Algorithm
,	O
orientation	O
preference	O
maps	O
,	O
pattern	O
learning/recognition	O
,	O
multi-target	O
tracking	O
,	O
image	O
stabilization	O
,	O
resolution	O
enhancement	O
,	O
image	O
deformations	O
and	O
mapping	O
,	O
image	O
inpainting	O
,	O
optical	O
flow	O
,	O
contouring	O
,	O
moving	B-Algorithm
object	I-Algorithm
detection	I-Algorithm
,	O
axis	O
of	O
symmetry	O
detection	O
,	O
and	O
image	B-General_Concept
fusion	I-General_Concept
.	O
</s>
<s>
Due	O
to	O
their	O
processing	O
capabilities	O
and	O
flexibility	O
,	O
CNN	B-Architecture
processors	O
have	O
been	O
used	O
and	O
prototyped	O
for	O
novel	O
field	O
applications	O
such	O
as	O
flame	O
analysis	O
for	O
monitoring	O
combustion	O
at	O
a	O
waste	O
incinerator	O
,	O
mine-detection	O
using	O
infrared	O
imagery	O
,	O
calorimeter	B-Algorithm
cluster	O
peak	O
for	O
high	O
energy	O
physics	O
,	O
anomaly	O
detection	O
in	O
potential	O
field	O
maps	O
for	O
geophysics	O
,	O
laser	O
dot	O
detection	O
,	O
metal	O
inspection	O
for	O
detecting	O
manufacturing	O
defects	O
,	O
and	O
seismic	O
horizon	O
picking	O
.	O
</s>
<s>
They	O
have	O
also	O
been	O
used	O
to	O
perform	O
biometric	O
functions	O
such	O
as	O
fingerprint	O
recognition	O
,	O
vein	O
feature	O
extraction	O
,	O
face	O
tracking	O
,	O
and	O
generating	O
visual	O
stimuli	O
via	O
emergent	O
patterns	O
to	O
gauge	O
perceptual	O
resonances	B-Application
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
been	O
used	O
for	O
medical	O
and	O
biological	O
research	O
in	O
performing	O
automated	O
nucleated	O
cell	O
counting	O
for	O
detecting	O
hyperplasia	O
,	O
segment	O
images	O
into	O
anatomically	O
and	O
pathologically	O
meaningful	O
regions	O
,	O
measure	O
and	O
quantify	O
cardiac	O
function	O
,	O
measure	O
the	O
timing	O
of	O
neurons	O
,	O
and	O
detect	O
brain	O
abnormalities	O
that	O
would	O
lead	O
to	O
seizures	O
.	O
</s>
<s>
One	O
potential	O
future	O
application	O
of	O
CNN	B-Architecture
microprocessors	O
is	O
to	O
combine	O
them	O
with	O
DNA	O
microarrays	O
to	O
allow	O
for	O
a	O
near-real	O
time	O
DNA	O
analysis	O
of	O
hundreds	O
of	O
thousands	O
of	O
different	O
DNA	O
sequences	O
.	O
</s>
<s>
Currently	O
,	O
the	O
major	O
bottleneck	O
of	O
DNA	O
microarray	O
analysis	O
is	O
the	O
amount	O
of	O
time	O
needed	O
to	O
process	O
data	O
in	O
the	O
form	O
of	O
images	O
,	O
and	O
using	O
a	O
CNN	B-Architecture
microprocessor	O
,	O
researchers	O
have	O
reduced	O
the	O
amount	O
of	O
time	O
needed	O
to	O
perform	O
this	O
calculation	O
to	O
7ms	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
also	O
been	O
used	O
to	O
generate	O
and	O
analyze	O
patterns	O
and	O
textures	O
.	O
</s>
<s>
One	O
motivation	O
was	O
to	O
use	O
CNN	B-Architecture
processors	O
to	O
understand	O
pattern	O
generation	O
in	O
natural	O
systems	O
.	O
</s>
<s>
Also	O
,	O
CNN	B-Architecture
processors	O
were	O
used	O
to	O
approximate	O
pattern	O
generation	O
systems	O
that	O
create	O
stationary	O
fronts	O
,	O
spatio-temporal	O
patterns	O
oscillating	B-Algorithm
in	O
time	O
,	O
hysteresis	O
,	O
memory	O
,	O
and	O
heterogeneity	O
.	O
</s>
<s>
There	O
is	O
an	O
ongoing	O
effort	O
to	O
incorporate	O
CNN	B-Architecture
processors	O
into	O
sensory-computing-actuating	O
machines	O
as	O
part	O
of	O
the	O
emerging	O
field	O
of	O
Cellular	O
Machines	O
.	O
</s>
<s>
The	O
basic	O
premise	O
is	O
to	O
create	O
an	O
integrated	O
system	O
that	O
uses	O
CNN	B-Architecture
processors	O
for	O
the	O
sensory	O
signal-processing	O
and	O
potentially	O
the	O
decision-making	O
and	O
control	O
.	O
</s>
<s>
The	O
reason	O
is	O
that	O
CNN	B-Architecture
processors	O
can	O
provide	O
a	O
low	O
power	O
,	O
small	O
size	O
,	O
and	O
eventually	O
low-cost	O
computing	O
and	O
actuating	O
system	O
suited	O
for	O
Cellular	O
Machines	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
been	O
proven	O
versatile	O
enough	O
for	O
some	O
control	O
functions	O
.	O
</s>
<s>
They	O
have	O
been	O
used	O
to	O
optimize	O
function	O
via	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
,	O
to	O
measure	O
distances	O
,	O
to	O
perform	O
optimal	O
path-finding	O
in	O
a	O
complex	O
,	O
dynamic	O
environment	O
,	O
and	O
theoretically	O
can	O
be	O
used	O
to	O
learn	O
and	O
associate	O
complex	O
stimuli	O
.	O
</s>
<s>
They	O
were	O
able	O
to	O
function	O
using	O
only	O
feedback	O
from	O
the	O
environment	O
,	O
allowing	O
for	O
a	O
robust	B-Application
,	O
flexible	O
,	O
biologically	O
inspired	O
robot	O
motor	O
system	O
.	O
</s>
<s>
CNN-based	O
systems	O
were	O
able	O
to	O
operate	O
in	O
different	O
environments	O
and	O
still	O
function	O
if	O
some	O
of	O
the	O
processing	O
units	O
are	O
disabled	O
.	O
</s>
<s>
The	O
variety	O
of	O
dynamical	O
behavior	O
seen	O
in	O
CNN	B-Architecture
processors	O
make	O
them	O
intriguing	O
for	O
communication	O
systems	O
.	O
</s>
<s>
Chaotic	O
communications	O
using	O
CNN	B-Architecture
processors	O
is	O
being	O
researched	O
due	O
to	O
their	O
potential	O
low	O
power	O
consumption	O
,	O
robustness	B-Application
and	O
spread	O
spectrum	O
features	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
can	O
be	O
used	O
on	O
both	O
the	O
transmitter	O
and	O
receiver	O
end	O
to	O
encode	O
and	O
decode	O
a	O
given	O
message	O
.	O
</s>
<s>
They	O
can	O
also	O
be	O
used	O
for	O
data	O
encryption	O
and	O
decryption	O
,	O
source	O
authentication	O
through	O
watermarking	O
,	O
detecting	O
of	O
complex	O
patterns	O
in	O
spectrogram	O
images	O
(	O
sound	B-Algorithm
processing	I-Algorithm
)	O
,	O
and	O
transient	O
spectral	O
signals	O
detection	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
are	O
neuromorphic	O
processors	O
,	O
meaning	O
that	O
they	O
emulate	O
certain	O
aspects	O
of	O
biological	B-General_Concept
neural	I-General_Concept
networks	I-General_Concept
.	O
</s>
<s>
The	O
original	O
CNN	B-Architecture
processors	O
were	O
based	O
on	O
mammalian	O
retinas	O
,	O
which	O
consist	O
of	O
a	O
layer	O
of	O
photodetectors	O
connected	O
to	O
several	O
layers	O
of	O
locally	O
coupled	O
neurons	O
.	O
</s>
<s>
This	O
makes	O
CNN	B-Architecture
processors	O
part	O
of	O
an	O
interdisciplinary	O
research	O
area	O
whose	O
goal	O
is	O
to	O
design	O
systems	O
that	O
leverage	O
knowledge	O
and	O
ideas	O
from	O
neuroscience	O
and	O
contribute	O
back	O
via	O
real-world	O
validation	O
of	O
theories	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
implemented	O
a	O
real-time	O
system	O
that	O
replicates	O
mammalian	O
retinas	O
,	O
validating	O
that	O
the	O
original	O
CNN	B-Architecture
architecture	O
chosen	O
modeled	O
the	O
correct	O
aspects	O
of	O
the	O
biological	B-General_Concept
neural	I-General_Concept
networks	I-General_Concept
used	O
to	O
perform	O
the	O
task	O
in	O
mammalian	O
life	O
.	O
</s>
<s>
However	O
,	O
CNN	B-Architecture
processors	O
are	O
not	O
limited	O
to	O
verifying	O
biological	B-General_Concept
neural	I-General_Concept
networks	I-General_Concept
associated	O
with	O
vision	O
processing	O
;	O
they	O
have	O
been	O
used	O
to	O
simulate	O
dynamic	O
activity	O
seen	O
in	O
mammalian	O
neural	B-General_Concept
networks	I-General_Concept
found	O
in	O
the	O
olfactory	O
bulb	O
and	O
locust	O
antennal	O
lobe	O
,	O
responsible	O
for	O
pre-processing	O
sensory	O
information	O
to	O
detect	O
differences	O
in	O
repeating	O
patterns	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
are	O
being	O
used	O
to	O
understand	O
systems	O
that	O
can	O
be	O
modeled	O
using	O
simple	O
,	O
coupled	O
units	O
,	O
such	O
as	O
living	O
cells	O
,	O
biological	O
networks	O
,	O
physiological	O
systems	O
,	O
and	O
ecosystems	O
.	O
</s>
<s>
The	O
CNN	B-Architecture
architecture	O
captures	O
some	O
of	O
the	O
dynamics	O
often	O
seen	O
in	O
nature	O
and	O
is	O
simple	O
enough	O
to	O
analyze	O
and	O
conduct	O
experiments	O
.	O
</s>
<s>
Instances	O
of	O
3d	O
CNN	B-Architecture
have	O
been	O
used	O
to	O
prove	O
certain	O
emergent	O
phenomena	O
in	O
complex	O
systems	O
,	O
establishing	O
a	O
link	O
between	O
art	O
,	O
dynamical	O
systems	O
and	O
VLSI	O
technology	O
.	O
</s>
<s>
CNN	B-Architecture
processors	O
have	O
been	O
used	O
to	O
research	O
a	O
variety	O
of	O
mathematical	O
concepts	O
,	O
such	O
as	O
non-equilibrium	O
systems	O
,	O
constructing	O
non-linear	O
systems	O
of	O
arbitrary	O
complexity	O
,	O
emergent	O
chaotic	O
dynamics	O
,	O
and	O
discovering	O
new	O
dynamic	O
behavior	O
.	O
</s>
<s>
Topics	O
explored	O
are	O
emergence	O
,	O
collective	O
behavior	O
,	O
local	O
activity	O
and	O
its	O
impact	O
on	O
global	O
behavior	O
,	O
and	O
quantifying	O
the	O
complexity	O
of	O
an	O
approximately	O
spatial	O
and	O
topologically	B-Architecture
invariant	O
system	O
.	O
</s>
