<s>
A	O
neural	B-Architecture
network	I-Architecture
can	O
refer	O
to	O
a	O
neural	B-General_Concept
circuit	I-General_Concept
of	O
biological	O
neurons	O
(	O
sometimes	O
also	O
called	O
a	O
biological	B-General_Concept
neural	I-General_Concept
network	I-General_Concept
)	O
,	O
a	O
network	O
of	O
artificial	B-Algorithm
neurons	I-Algorithm
or	O
nodes	O
in	O
the	O
case	O
of	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
.	O
</s>
<s>
Artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
are	O
used	O
for	O
solving	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
problems	O
;	O
they	O
model	O
connections	O
of	O
biological	O
neurons	O
as	O
weights	O
between	O
nodes	O
.	O
</s>
<s>
Finally	O
,	O
an	O
activation	O
function	O
controls	O
the	O
amplitude	B-Application
of	O
the	O
output	O
.	O
</s>
<s>
These	O
artificial	O
networks	O
may	O
be	O
used	O
for	O
predictive	B-General_Concept
modeling	I-General_Concept
,	O
adaptive	O
control	O
and	O
applications	O
where	O
they	O
can	O
be	O
trained	O
via	O
a	O
dataset	O
.	O
</s>
<s>
A	O
biological	B-General_Concept
neural	I-General_Concept
network	I-General_Concept
is	O
composed	O
of	O
a	O
group	O
of	O
chemically	O
connected	O
or	O
functionally	O
associated	O
neurons	O
.	O
</s>
<s>
Connections	O
,	O
called	O
synapses	B-Application
,	O
are	O
usually	O
formed	O
from	O
axons	B-Algorithm
to	O
dendrites	O
,	O
though	O
dendrodendritic	O
synapses	B-Application
and	O
other	O
connections	O
are	O
possible	O
.	O
</s>
<s>
Artificial	B-Application
intelligence	I-Application
,	O
cognitive	O
modelling	O
,	O
and	O
neural	B-Architecture
networks	I-Architecture
are	O
information	O
processing	O
paradigms	O
inspired	O
by	O
how	O
biological	O
neural	O
systems	O
process	O
data	O
.	O
</s>
<s>
Artificial	B-Application
intelligence	I-Application
and	O
cognitive	O
modelling	O
try	O
to	O
simulate	O
some	O
properties	O
of	O
biological	B-General_Concept
neural	I-General_Concept
networks	I-General_Concept
.	O
</s>
<s>
In	O
the	B-Application
artificial	I-Application
intelligence	I-Application
field	O
,	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
have	O
been	O
applied	O
successfully	O
to	O
speech	B-Application
recognition	I-Application
,	O
image	B-General_Concept
analysis	I-General_Concept
and	O
adaptive	O
control	O
,	O
in	O
order	O
to	O
construct	O
software	B-General_Concept
agents	I-General_Concept
(	O
in	O
computer	O
and	O
video	O
games	O
)	O
or	O
autonomous	O
robots	O
.	O
</s>
<s>
Historically	O
,	O
digital	B-General_Concept
computers	O
evolved	O
from	O
the	O
von	B-Architecture
Neumann	I-Architecture
model	I-Architecture
,	O
and	O
operate	O
via	O
the	O
execution	O
of	O
explicit	O
instructions	O
via	O
access	O
to	O
memory	B-Architecture
by	O
a	O
number	O
of	O
processors	O
.	O
</s>
<s>
On	O
the	O
other	O
hand	O
,	O
the	O
origins	O
of	O
neural	B-Architecture
networks	I-Architecture
are	O
based	O
on	O
efforts	O
to	O
model	O
information	O
processing	O
in	O
biological	O
systems	O
.	O
</s>
<s>
Unlike	O
the	O
von	B-Architecture
Neumann	I-Architecture
model	I-Architecture
,	O
neural	B-Architecture
network	I-Architecture
computing	O
does	O
not	O
separate	O
memory	B-Architecture
and	O
processing	O
.	O
</s>
<s>
Neural	B-Architecture
network	I-Architecture
theory	O
has	O
served	O
to	O
identify	O
better	O
how	O
the	O
neurons	O
in	O
the	O
brain	O
function	O
and	O
provide	O
the	O
basis	O
for	O
efforts	O
to	O
create	O
artificial	B-Application
intelligence	I-Application
.	I-Application
</s>
<s>
The	O
preliminary	O
theoretical	O
base	O
for	O
contemporary	O
neural	B-Architecture
networks	I-Architecture
was	O
independently	O
proposed	O
by	O
Alexander	O
Bain	O
(	O
1873	O
)	O
and	O
William	O
James	O
(	O
1890	O
)	O
.	O
</s>
<s>
According	O
to	O
his	O
theory	O
,	O
this	O
repetition	O
was	O
what	O
led	O
to	O
the	O
formation	O
of	O
memory	B-Architecture
.	O
</s>
<s>
His	O
model	O
,	O
by	O
focusing	O
on	O
the	O
flow	O
of	O
electrical	O
currents	O
,	O
did	O
not	O
require	O
individual	O
neural	O
connections	O
for	O
each	O
memory	B-Architecture
or	O
action	O
.	O
</s>
<s>
Wilhelm	O
Lenz	O
(	O
1920	O
)	O
and	O
Ernst	O
Ising	O
(	O
1925	O
)	O
created	O
and	O
analyzed	O
the	O
Ising	O
model	O
which	O
is	O
essentially	O
a	O
non-learning	O
artificial	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
(	O
RNN	O
)	O
consisting	O
of	O
neuron-like	O
threshold	O
elements	O
.	O
</s>
<s>
McCulloch	O
and	O
Pitts	O
(	O
1943	O
)	O
also	O
created	O
a	O
computational	O
model	O
for	O
neural	B-Architecture
networks	I-Architecture
based	O
on	O
mathematics	O
and	O
algorithms	O
.	O
</s>
<s>
These	O
early	O
models	O
paved	O
the	O
way	O
for	O
neural	B-Architecture
network	I-Architecture
research	O
to	O
split	O
into	O
two	O
distinct	O
approaches	O
.	O
</s>
<s>
One	O
approach	O
focused	O
on	O
biological	O
processes	O
in	O
the	O
brain	O
and	O
the	O
other	O
focused	O
on	O
the	O
application	O
of	O
neural	B-Architecture
networks	I-Architecture
to	O
artificial	B-Application
intelligence	I-Application
.	I-Application
</s>
<s>
Hebbian	O
learning	O
is	O
considered	O
to	O
be	O
a	O
'	O
typical	O
 '	O
unsupervised	B-General_Concept
learning	I-General_Concept
rule	O
and	O
its	O
later	O
variants	O
were	O
early	O
models	O
for	O
long	O
term	O
potentiation	O
.	O
</s>
<s>
Other	O
neural	B-Architecture
network	I-Architecture
computational	O
machines	O
were	O
created	O
by	O
Rochester	O
,	O
Holland	O
,	O
Habit	O
,	O
and	O
Duda	O
(	O
1956	O
)	O
.	O
</s>
<s>
Frank	O
Rosenblatt	O
(	O
1958	O
)	O
created	O
the	O
perceptron	B-Algorithm
,	O
an	O
algorithm	O
for	O
pattern	O
recognition	O
based	O
on	O
a	O
two-layer	O
learning	O
computer	O
network	O
using	O
simple	O
addition	O
and	O
subtraction	O
.	O
</s>
<s>
With	O
mathematical	O
notation	O
,	O
Rosenblatt	O
also	O
described	O
circuitry	O
not	O
in	O
the	O
basic	O
perceptron	B-Algorithm
,	O
such	O
as	O
the	O
exclusive-or	O
circuit	O
.	O
</s>
<s>
Some	O
say	O
that	O
neural	B-Architecture
network	I-Architecture
research	O
stagnated	O
after	O
the	O
publication	O
of	O
machine	O
learning	O
research	O
by	O
Marvin	O
Minsky	O
and	O
Seymour	O
Papert	O
(	O
1969	O
)	O
.	O
</s>
<s>
They	O
discovered	O
two	O
key	O
issues	O
with	O
the	O
computational	O
machines	O
that	O
processed	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
The	O
first	O
issue	O
was	O
that	O
single-layer	O
neural	B-Architecture
networks	I-Architecture
were	O
incapable	O
of	O
processing	O
the	O
exclusive-or	O
circuit	O
.	O
</s>
<s>
The	O
second	O
significant	O
issue	O
was	O
that	O
computers	O
were	O
not	O
sophisticated	O
enough	O
to	O
effectively	O
handle	O
the	O
long	O
run	O
time	O
required	O
by	O
large	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
However	O
,	O
by	O
the	O
time	O
this	O
book	O
came	O
out	O
,	O
methods	O
for	O
training	O
multilayer	B-Algorithm
perceptrons	I-Algorithm
(	O
MLPs	O
)	O
were	O
already	O
known	O
.	O
</s>
<s>
The	O
first	O
deep	B-Algorithm
learning	I-Algorithm
MLP	O
was	O
published	O
by	O
Alexey	O
Grigorevich	O
Ivakhnenko	O
and	O
Valentin	O
Lapa	O
in	O
1965	O
.	O
</s>
<s>
The	O
first	O
deep	B-Algorithm
learning	I-Algorithm
MLP	O
trained	O
by	O
stochastic	B-Algorithm
gradient	I-Algorithm
descent	I-Algorithm
was	O
published	O
in	O
1967	O
by	O
Shun'ichi	O
Amari	O
.	O
</s>
<s>
Neural	B-Architecture
network	I-Architecture
research	O
was	O
boosted	O
when	O
computers	O
achieved	O
greater	O
processing	O
power	O
.	O
</s>
<s>
Also	O
key	O
in	O
later	O
advances	O
was	O
the	O
backpropagation	B-Algorithm
algorithm	O
.	O
</s>
<s>
the	O
reverse	O
mode	O
of	O
automatic	B-Algorithm
differentiation	I-Algorithm
or	O
reverse	O
accumulation	O
,	O
due	O
to	O
Seppo	O
Linnainmaa	O
(	O
1970	O
)	O
.	O
</s>
<s>
The	O
term	O
"	O
back-propagating	O
errors	O
"	O
was	O
introduced	O
in	O
1962	O
by	O
Frank	O
Rosenblatt	O
,	O
but	O
he	O
did	O
not	O
have	O
an	O
implementation	O
of	O
this	O
procedure	O
,	O
although	O
Henry	O
J	O
.	O
Kelley	O
had	O
a	O
continuous	O
precursor	O
of	O
backpropagation	B-Algorithm
already	O
in	O
1960	O
in	O
the	O
context	O
of	O
control	O
theory	O
.	O
</s>
<s>
In	O
1982	O
,	O
Paul	O
Werbos	O
applied	O
backpropagation	B-Algorithm
to	O
MLPs	O
in	O
the	O
way	O
that	O
has	O
become	O
standard	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
,	O
as	O
used	O
in	O
artificial	B-Application
intelligence	I-Application
,	O
have	O
traditionally	O
been	O
viewed	O
as	O
simplified	O
models	O
of	O
neural	O
processing	O
in	O
the	O
brain	O
,	O
even	O
though	O
the	O
relation	O
between	O
this	O
model	O
and	O
brain	O
biological	O
architecture	O
is	O
debated	O
,	O
as	O
it	O
is	O
not	O
clear	O
to	O
what	O
degree	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
mirror	O
brain	O
function	O
.	O
</s>
<s>
A	O
neural	B-Architecture
network	I-Architecture
(	O
NN	O
)	O
,	O
in	O
the	O
case	O
of	O
artificial	B-Algorithm
neurons	I-Algorithm
called	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
ANN	O
)	O
or	O
simulated	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
SNN	O
)	O
,	O
is	O
an	O
interconnected	O
group	O
of	O
natural	O
or	O
artificial	B-Algorithm
neurons	I-Algorithm
that	O
uses	O
a	O
mathematical	O
or	O
computational	O
model	O
for	O
information	O
processing	O
based	O
on	O
a	O
connectionistic	O
approach	O
to	O
computation	O
.	O
</s>
<s>
In	O
more	O
practical	O
terms	O
neural	B-Architecture
networks	I-Architecture
are	O
non-linear	O
statistical	O
data	B-Application
modeling	I-Application
or	O
decision	O
making	O
tools	O
.	O
</s>
<s>
An	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
involves	O
a	O
network	O
of	O
simple	O
processing	O
elements	O
(	O
artificial	B-Algorithm
neurons	I-Algorithm
)	O
which	O
can	O
exhibit	O
complex	O
global	O
behavior	O
,	O
determined	O
by	O
the	O
connections	O
between	O
the	O
processing	O
elements	O
and	O
element	O
parameters	O
.	O
</s>
<s>
Artificial	B-Algorithm
neurons	I-Algorithm
were	O
first	O
proposed	O
in	O
1943	O
by	O
Warren	O
McCulloch	O
,	O
a	O
neurophysiologist	O
,	O
and	O
Walter	O
Pitts	O
,	O
a	O
logician	O
,	O
who	O
first	O
collaborated	O
at	O
the	O
University	O
of	O
Chicago	O
.	O
</s>
<s>
One	O
classical	O
type	O
of	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
is	O
the	O
recurrent	B-Algorithm
Hopfield	B-Algorithm
network	I-Algorithm
.	O
</s>
<s>
The	O
concept	O
of	O
a	O
neural	B-Architecture
network	I-Architecture
appears	O
to	O
have	O
first	O
been	O
proposed	O
by	O
Alan	O
Turing	O
in	O
his	O
1948	O
paper	O
Intelligent	O
Machinery	O
in	O
which	O
he	O
called	O
them	O
"	O
B-type	O
unorganised	O
machines	O
"	O
.	O
</s>
<s>
The	O
utility	O
of	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
models	O
lies	O
in	O
the	O
fact	O
that	O
they	O
can	O
be	O
used	O
to	O
infer	O
a	O
function	O
from	O
observations	O
and	O
also	O
to	O
use	O
it	O
.	O
</s>
<s>
Unsupervised	O
neural	B-Architecture
networks	I-Architecture
can	O
also	O
be	O
used	O
to	O
learn	O
representations	O
of	O
the	O
input	O
that	O
capture	O
the	O
salient	O
characteristics	O
of	O
the	O
input	O
distribution	O
,	O
e.g.	O
,	O
see	O
the	O
Boltzmann	B-Algorithm
machine	I-Algorithm
(	O
1983	O
)	O
,	O
and	O
more	O
recently	O
,	O
deep	B-Algorithm
learning	I-Algorithm
algorithms	O
,	O
which	O
can	O
implicitly	O
learn	O
the	O
distribution	O
function	O
of	O
the	O
observed	O
data	O
.	O
</s>
<s>
Learning	O
in	O
neural	B-Architecture
networks	I-Architecture
is	O
particularly	O
useful	O
in	O
applications	O
where	O
the	O
complexity	O
of	O
the	O
data	O
or	O
task	O
makes	O
the	O
design	O
of	O
such	O
functions	O
by	O
hand	O
impractical	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
can	O
be	O
used	O
in	O
different	O
fields	O
.	O
</s>
<s>
The	O
tasks	O
to	O
which	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
are	O
applied	O
tend	O
to	O
fall	O
within	O
the	O
following	O
broad	O
categories	O
:	O
</s>
<s>
Classification	B-General_Concept
,	O
including	O
pattern	O
and	O
sequence	O
recognition	O
,	O
novelty	B-Algorithm
detection	I-Algorithm
and	O
sequential	O
decision	O
making	O
.	O
</s>
<s>
Data	B-General_Concept
processing	I-General_Concept
,	O
including	O
filtering	O
,	O
clustering	O
,	O
blind	B-Application
signal	I-Application
separation	I-Application
and	O
compression	B-General_Concept
.	O
</s>
<s>
Application	O
areas	O
of	O
ANNs	O
include	O
nonlinear	O
system	O
identification	O
and	O
control	O
(	O
vehicle	O
control	O
,	O
process	O
control	O
)	O
,	O
game-playing	O
and	O
decision	O
making	O
(	O
backgammon	O
,	O
chess	O
,	O
racing	O
)	O
,	O
pattern	O
recognition	O
(	O
radar	O
systems	O
,	O
face	O
identification	O
,	O
object	O
recognition	O
)	O
,	O
sequence	O
recognition	O
(	O
gesture	O
,	O
speech	O
,	O
handwritten	B-Application
text	I-Application
recognition	I-Application
)	O
,	O
medical	O
diagnosis	O
,	O
financial	O
applications	O
,	O
data	B-Application
mining	I-Application
(	O
or	O
knowledge	B-Application
discovery	I-Application
in	I-Application
databases	I-Application
,	O
"	O
KDD	O
"	O
)	O
,	O
visualization	O
and	O
e-mail	O
spam	O
filtering	O
.	O
</s>
<s>
Theoretical	O
and	O
computational	O
neuroscience	O
is	O
the	O
field	O
concerned	O
with	O
the	O
analysis	O
and	O
computational	B-Application
modeling	I-Application
of	O
biological	O
neural	O
systems	O
.	O
</s>
<s>
To	O
gain	O
this	O
understanding	O
,	O
neuroscientists	O
strive	O
to	O
make	O
a	O
link	O
between	O
observed	O
biological	O
processes	O
(	O
data	O
)	O
,	O
biologically	O
plausible	O
mechanisms	O
for	O
neural	O
processing	O
and	O
learning	O
(	O
biological	B-General_Concept
neural	I-General_Concept
network	I-General_Concept
models	O
)	O
and	O
theory	O
(	O
statistical	O
learning	O
theory	O
and	O
information	O
theory	O
)	O
.	O
</s>
<s>
They	O
range	O
from	O
models	O
of	O
the	O
short-term	O
behaviour	O
of	O
individual	O
neurons	O
,	O
through	O
models	O
of	O
the	O
dynamics	O
of	O
neural	B-Architecture
circuitry	I-Architecture
arising	O
from	O
interactions	O
between	O
individual	O
neurons	O
,	O
to	O
models	O
of	O
behaviour	O
arising	O
from	O
abstract	O
neural	O
modules	O
that	O
represent	O
complete	O
subsystems	O
.	O
</s>
<s>
These	O
include	O
models	O
of	O
the	O
long-term	O
and	O
short-term	O
plasticity	O
of	O
neural	O
systems	O
and	O
its	O
relation	O
to	O
learning	O
and	O
memory	B-Architecture
,	O
from	O
the	O
individual	O
neuron	O
to	O
the	O
system	O
level	O
.	O
</s>
<s>
In	O
August	O
2020	O
scientists	O
reported	O
that	O
bi-directional	O
connections	O
,	O
or	O
added	O
appropriate	O
feedback	O
connections	O
,	O
can	O
accelerate	O
and	O
improve	O
communication	O
between	O
and	O
in	O
modular	O
neural	B-Architecture
networks	I-Architecture
of	O
the	O
brain	O
's	O
cerebral	O
cortex	O
and	O
lower	O
the	O
threshold	O
for	O
their	O
successful	O
communication	O
.	O
</s>
<s>
Historically	O
,	O
a	O
common	O
criticism	O
of	O
neural	B-Architecture
networks	I-Architecture
,	O
particularly	O
in	O
robotics	O
,	O
was	O
that	O
they	O
require	O
a	O
large	O
diversity	O
of	O
training	O
samples	O
for	O
real-world	O
operation	O
.	O
</s>
<s>
Dean	O
Pomerleau	O
,	O
in	O
his	O
research	O
presented	O
in	O
the	O
paper	O
"	O
Knowledge-based	O
Training	O
of	O
Artificial	B-Architecture
Neural	I-Architecture
Networks	I-Architecture
for	O
Autonomous	O
Robot	O
Driving	O
,	O
"	O
uses	O
a	O
neural	B-Architecture
network	I-Architecture
to	O
train	O
a	O
robotic	O
vehicle	O
to	O
drive	O
on	O
multiple	O
types	O
of	O
roads	O
(	O
single	O
lane	O
,	O
multi-lane	O
,	O
dirt	O
,	O
etc	O
.	O
)	O
.	O
</s>
<s>
These	O
issues	O
are	O
common	O
in	O
neural	B-Architecture
networks	I-Architecture
that	O
must	O
decide	O
from	O
amongst	O
a	O
wide	O
variety	O
of	O
responses	O
,	O
but	O
can	O
be	O
dealt	O
with	O
in	O
several	O
ways	O
,	O
for	O
example	O
by	O
randomly	O
shuffling	O
the	O
training	O
examples	O
,	O
by	O
using	O
a	O
numerical	O
optimization	O
algorithm	O
that	O
does	O
not	O
take	O
too	O
large	O
steps	O
when	O
changing	O
the	O
network	O
connections	O
following	O
an	O
example	O
,	O
or	O
by	O
grouping	O
examples	O
in	O
so-called	O
mini-batches	O
.	O
</s>
<s>
A	O
.	O
K	O
.	O
Dewdney	O
,	O
a	O
former	O
Scientific	O
American	O
columnist	O
,	O
wrote	O
in	O
1997	O
,	O
"	O
Although	O
neural	B-Architecture
nets	I-Architecture
do	O
solve	O
a	O
few	O
toy	O
problems	O
,	O
their	O
powers	O
of	O
computation	O
are	O
so	O
limited	O
that	O
I	O
am	O
surprised	O
anyone	O
takes	O
them	O
seriously	O
as	O
a	O
general	O
problem-solving	O
tool.	O
"	O
</s>
<s>
Arguments	O
for	O
Dewdney	O
's	O
position	O
are	O
that	O
to	O
implement	O
large	O
and	O
effective	O
software	O
neural	B-Architecture
networks	I-Architecture
,	O
much	O
processing	O
and	O
storage	O
resources	O
need	O
to	O
be	O
committed	O
.	O
</s>
<s>
While	O
the	O
brain	O
has	O
hardware	O
tailored	O
to	O
the	O
task	O
of	O
processing	O
signals	O
through	O
a	O
graph	O
of	O
neurons	O
,	O
simulating	O
even	O
a	O
most	O
simplified	O
form	O
on	O
Von	O
Neumann	O
technology	O
may	O
compel	O
a	O
neural	B-Architecture
network	I-Architecture
designer	O
to	O
fill	O
many	O
millions	O
of	O
database	O
rows	O
for	O
its	O
connections	O
—	O
which	O
can	O
consume	O
vast	O
amounts	O
of	O
computer	O
memory	B-Architecture
and	O
data	B-General_Concept
storage	I-General_Concept
capacity	O
.	O
</s>
<s>
Furthermore	O
,	O
the	O
designer	O
of	O
neural	B-Architecture
network	I-Architecture
systems	O
will	O
often	O
need	O
to	O
simulate	O
the	O
transmission	O
of	O
signals	O
through	O
many	O
of	O
these	O
connections	O
and	O
their	O
associated	O
neurons	O
—	O
which	O
must	O
often	O
be	O
matched	O
with	O
incredible	O
amounts	O
of	O
CPU	B-Device
processing	O
power	O
and	O
time	O
.	O
</s>
<s>
While	O
neural	B-Architecture
networks	I-Architecture
often	O
yield	O
effective	O
programs	O
,	O
they	O
too	O
often	O
do	O
so	O
at	O
the	O
cost	O
of	O
efficiency	O
(	O
they	O
tend	O
to	O
consume	O
considerable	O
amounts	O
of	O
time	O
and	O
money	O
)	O
.	O
</s>
<s>
Arguments	O
against	O
Dewdney	O
's	O
position	O
are	O
that	O
neural	B-Architecture
nets	I-Architecture
have	O
been	O
successfully	O
used	O
to	O
solve	O
many	O
complex	O
and	O
diverse	O
tasks	O
,	O
such	O
as	O
autonomously	O
flying	O
aircraft	O
.	O
</s>
<s>
Technology	O
writer	O
Roger	O
Bridgman	O
commented	O
on	O
Dewdney	O
's	O
statements	O
about	O
neural	B-Architecture
nets	I-Architecture
:	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
,	O
for	O
instance	O
,	O
are	O
in	O
the	O
dock	O
not	O
only	O
because	O
they	O
have	O
been	O
hyped	O
to	O
high	O
heaven	O
,	O
(	O
what	O
has	O
n't	O
?	O
)	O
</s>
<s>
In	O
spite	O
of	O
his	O
emphatic	O
declaration	O
that	O
science	O
is	O
not	O
technology	O
,	O
Dewdney	O
seems	O
here	O
to	O
pillory	O
neural	B-Architecture
nets	I-Architecture
as	O
bad	O
science	O
when	O
most	O
of	O
those	O
devising	O
them	O
are	O
just	O
trying	O
to	O
be	O
good	O
engineers	O
.	O
</s>
<s>
Although	O
it	O
is	O
true	O
that	O
analyzing	O
what	O
has	O
been	O
learned	O
by	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
is	O
difficult	O
,	O
it	O
is	O
much	O
easier	O
to	O
do	O
so	O
than	O
to	O
analyze	O
what	O
has	O
been	O
learned	O
by	O
a	O
biological	B-General_Concept
neural	I-General_Concept
network	I-General_Concept
.	O
</s>
<s>
Moreover	O
,	O
recent	O
emphasis	O
on	O
the	O
explainability	O
of	O
AI	B-Application
has	O
contributed	O
towards	O
the	O
development	O
of	O
methods	O
,	O
notably	O
those	O
based	O
on	O
attention	O
mechanisms	O
,	O
for	O
visualizing	O
and	O
explaining	O
learned	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
Furthermore	O
,	O
researchers	O
involved	O
in	O
exploring	O
learning	O
algorithms	O
for	O
neural	B-Architecture
networks	I-Architecture
are	O
gradually	O
uncovering	O
generic	O
principles	O
that	O
allow	O
a	O
learning	O
machine	O
to	O
be	O
successful	O
.	O
</s>
<s>
Some	O
other	O
criticisms	O
came	O
from	O
believers	O
of	O
hybrid	O
models	O
(	O
combining	O
neural	B-Architecture
networks	I-Architecture
and	O
symbolic	B-General_Concept
approaches	O
)	O
.	O
</s>
<s>
Research	O
is	O
ongoing	O
in	O
understanding	O
the	O
computational	O
algorithms	O
used	O
in	O
the	O
brain	O
,	O
with	O
some	O
recent	O
biological	O
evidence	O
for	O
radial	B-Algorithm
basis	I-Algorithm
networks	I-Algorithm
and	O
neural	O
backpropagation	B-Algorithm
as	O
mechanisms	O
for	O
processing	O
data	O
.	O
</s>
<s>
More	O
recent	O
efforts	O
show	O
promise	O
for	O
creating	O
nanodevices	O
for	O
very	O
large	O
scale	O
principal	B-Application
components	I-Application
analyses	O
and	O
convolution	B-Language
.	O
</s>
<s>
If	O
successful	O
,	O
these	O
efforts	O
could	O
usher	O
in	O
a	O
new	O
era	O
of	O
neural	B-Architecture
computing	I-Architecture
that	O
is	O
a	O
step	O
beyond	O
digital	B-General_Concept
computing	O
,	O
because	O
it	O
depends	O
on	O
learning	O
rather	O
than	O
programming	O
and	O
because	O
it	O
is	O
fundamentally	O
analog	O
rather	O
than	O
digital	B-General_Concept
even	O
though	O
the	O
first	O
instantiations	O
may	O
in	O
fact	O
be	O
with	O
CMOS	O
digital	B-General_Concept
devices	O
.	O
</s>
<s>
Between	O
2009	O
and	O
2012	O
,	O
the	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
and	O
deep	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
developed	O
in	O
the	O
research	O
group	O
of	O
Jürgen	O
Schmidhuber	O
at	O
the	O
Swiss	O
AI	B-Application
Lab	O
IDSIA	O
have	O
won	O
eight	O
international	O
competitions	O
in	O
pattern	O
recognition	O
and	O
machine	O
learning	O
.	O
</s>
<s>
For	O
example	O
,	O
multi-dimensional	O
long	B-Algorithm
short	I-Algorithm
term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
won	O
three	O
competitions	O
in	O
connected	O
handwriting	B-Application
recognition	I-Application
at	O
the	O
2009	O
International	O
Conference	O
on	O
Document	O
Analysis	O
and	O
Recognition	O
(	O
ICDAR	O
)	O
,	O
without	O
any	O
prior	O
knowledge	O
about	O
the	O
three	O
different	O
languages	O
to	O
be	O
learned	O
.	O
</s>
<s>
Variants	O
of	O
the	O
back-propagation	B-Algorithm
algorithm	O
as	O
well	O
as	O
unsupervised	O
methods	O
by	O
Geoff	O
Hinton	O
and	O
colleagues	O
at	O
the	O
University	O
of	O
Toronto	O
can	O
be	O
used	O
to	O
train	O
deep	O
,	O
highly	O
nonlinear	O
neural	O
architectures	O
,	O
similar	O
to	O
the	O
1980	O
Neocognitron	B-Algorithm
by	O
Kunihiko	O
Fukushima	O
,	O
and	O
the	O
"	O
standard	O
architecture	O
of	O
vision	O
"	O
,	O
inspired	O
by	O
the	O
simple	O
and	O
complex	O
cells	O
identified	O
by	O
David	O
H	O
.	O
Hubel	O
and	O
Torsten	O
Wiesel	O
in	O
the	O
primary	O
visual	O
cortex	O
.	O
</s>
<s>
These	O
can	O
be	O
shown	O
to	O
offer	O
best	O
approximation	O
properties	O
and	O
have	O
been	O
applied	O
in	O
nonlinear	O
system	O
identification	O
and	O
classification	B-General_Concept
applications	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
feedforward	B-Algorithm
networks	O
alternate	O
convolutional	O
layers	O
and	O
max-pooling	O
layers	O
,	O
topped	O
by	O
several	O
pure	O
classification	B-General_Concept
layers	O
.	O
</s>
<s>
Fast	O
GPU-based	O
implementations	O
of	O
this	O
approach	O
have	O
won	O
several	O
pattern	O
recognition	O
contests	O
,	O
including	O
the	O
IJCNN	O
2011	O
Traffic	O
Sign	O
Recognition	O
Competition	O
and	O
the	O
ISBI	O
2012	O
Segmentation	O
of	O
Neuronal	O
Structures	O
in	O
Electron	O
Microscopy	O
Stacks	O
challenge	O
.	O
</s>
<s>
Such	O
neural	B-Architecture
networks	I-Architecture
also	O
were	O
the	O
first	O
artificial	O
pattern	O
recognizers	O
to	O
achieve	O
human-competitive	O
or	O
even	O
superhuman	O
performance	O
on	O
benchmarks	O
such	O
as	O
traffic	O
sign	O
recognition	O
(	O
IJCNN	O
2012	O
)	O
,	O
or	O
the	O
MNIST	O
handwritten	O
digits	O
problem	O
of	O
Yann	O
LeCun	O
and	O
colleagues	O
at	O
NYU	O
.	O
</s>
<s>
Analytical	O
and	O
computational	O
techniques	O
derived	O
from	O
statistical	O
physics	O
of	O
disordered	O
systems	O
,	O
can	O
be	O
extended	O
to	O
large-scale	O
problems	O
,	O
including	O
machine	O
learning	O
,	O
e.g.	O
,	O
to	O
analyze	O
the	O
weight	O
space	O
of	O
deep	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
