<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
part	O
of	O
a	O
broader	O
family	O
of	O
machine	O
learning	O
methods	O
,	O
which	O
is	O
based	O
on	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
with	O
representation	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
Learning	O
can	O
be	O
supervised	B-General_Concept
,	O
semi-supervised	B-General_Concept
or	O
unsupervised	B-General_Concept
.	O
</s>
<s>
Deep-learning	B-Algorithm
architectures	O
such	O
as	O
deep	O
neural	B-Architecture
networks	I-Architecture
,	O
deep	B-Algorithm
belief	I-Algorithm
networks	I-Algorithm
,	O
deep	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
,	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
and	O
transformers	B-Algorithm
have	O
been	O
applied	O
to	O
fields	O
including	O
computer	B-Application
vision	I-Application
,	O
speech	B-Application
recognition	I-Application
,	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
machine	B-Application
translation	I-Application
,	O
bioinformatics	O
,	O
drug	O
design	O
,	O
medical	B-Algorithm
image	I-Algorithm
analysis	I-Algorithm
,	O
climate	O
science	O
,	O
material	O
inspection	O
and	O
board	O
game	O
programs	O
,	O
where	O
they	O
have	O
produced	O
results	O
comparable	O
to	O
and	O
in	O
some	O
cases	O
surpassing	O
human	O
expert	O
performance	O
.	O
</s>
<s>
Artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
ANNs	O
)	O
were	O
inspired	O
by	O
information	O
processing	O
and	O
distributed	O
communication	O
nodes	O
in	O
biological	O
systems	O
.	O
</s>
<s>
Specifically	O
,	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
tend	O
to	O
be	O
static	O
and	O
symbolic	O
,	O
while	O
the	O
biological	O
brain	O
of	O
most	O
living	O
organisms	O
is	O
dynamic	O
(	O
plastic	O
)	O
and	O
analog	O
.	O
</s>
<s>
The	O
adjective	O
"	O
deep	O
"	O
in	O
deep	B-Algorithm
learning	I-Algorithm
refers	O
to	O
the	O
use	O
of	O
multiple	O
layers	O
in	O
the	O
network	O
.	O
</s>
<s>
Early	O
work	O
showed	O
that	O
a	O
linear	B-Algorithm
perceptron	I-Algorithm
cannot	O
be	O
a	O
universal	O
classifier	O
,	O
but	O
that	O
a	O
network	O
with	O
a	O
nonpolynomial	O
activation	B-Algorithm
function	I-Algorithm
with	O
one	O
hidden	O
layer	O
of	O
unbounded	O
width	O
can	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
a	O
modern	O
variation	O
that	O
is	O
concerned	O
with	O
an	O
unbounded	O
number	O
of	O
layers	O
of	O
bounded	O
size	O
,	O
which	O
permits	O
practical	O
application	O
and	O
optimized	O
implementation	O
,	O
while	O
retaining	O
theoretical	O
universality	O
under	O
mild	O
conditions	O
.	O
</s>
<s>
In	O
deep	B-Algorithm
learning	I-Algorithm
the	O
layers	O
are	O
also	O
permitted	O
to	O
be	O
heterogeneous	O
and	O
to	O
deviate	O
widely	O
from	O
biologically	O
informed	O
connectionist	O
models	O
,	O
for	O
the	O
sake	O
of	O
efficiency	O
,	O
trainability	O
and	O
understandability	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
a	O
class	O
of	O
machine	O
learning	O
algorithms	O
that	O
uses	O
multiple	O
layers	O
to	O
progressively	O
extract	O
higher-level	O
features	O
from	O
the	O
raw	O
input	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
image	B-Algorithm
processing	I-Algorithm
,	O
lower	O
layers	O
may	O
identify	O
edges	O
,	O
while	O
higher	O
layers	O
may	O
identify	O
the	O
concepts	O
relevant	O
to	O
a	O
human	O
such	O
as	O
digits	O
or	O
letters	O
or	O
faces	O
.	O
</s>
<s>
From	O
another	O
angle	O
to	O
view	O
deep	B-Algorithm
learning	I-Algorithm
,	O
deep	B-Algorithm
learning	I-Algorithm
refers	O
to	O
‘	O
computer-simulate	O
’	O
or	O
‘	O
automate’	O
human	O
learning	O
processes	O
from	O
a	O
source	O
(	O
e.g.	O
,	O
an	O
image	O
of	O
dogs	O
)	O
to	O
a	O
learned	O
object	O
(	O
dogs	O
)	O
.	O
</s>
<s>
Most	O
modern	O
deep	B-Algorithm
learning	I-Algorithm
models	O
are	O
based	O
on	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
,	O
specifically	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
CNN	B-Architecture
)	O
s	O
,	O
although	O
they	O
can	O
also	O
include	O
propositional	O
formulas	O
or	O
latent	O
variables	O
organized	O
layer-wise	O
in	O
deep	O
generative	O
models	O
such	O
as	O
the	O
nodes	O
in	O
deep	B-Algorithm
belief	I-Algorithm
networks	I-Algorithm
and	O
deep	O
Boltzmann	B-Algorithm
machines	I-Algorithm
.	O
</s>
<s>
In	O
deep	B-Algorithm
learning	I-Algorithm
,	O
each	O
level	O
learns	O
to	O
transform	O
its	O
input	O
data	O
into	O
a	O
slightly	O
more	O
abstract	O
and	O
composite	O
representation	O
.	O
</s>
<s>
In	O
an	O
image	O
recognition	O
application	O
,	O
the	O
raw	O
input	O
may	O
be	O
a	O
matrix	B-Architecture
of	O
pixels	O
;	O
the	O
first	O
representational	O
layer	O
may	O
abstract	O
the	O
pixels	O
and	O
encode	O
edges	O
;	O
the	O
second	O
layer	O
may	O
compose	O
and	O
encode	O
arrangements	O
of	O
edges	O
;	O
the	O
third	O
layer	O
may	O
encode	O
a	O
nose	O
and	O
eyes	O
;	O
and	O
the	O
fourth	O
layer	O
may	O
recognize	O
that	O
the	O
image	O
contains	O
a	O
face	O
.	O
</s>
<s>
Importantly	O
,	O
a	O
deep	B-Algorithm
learning	I-Algorithm
process	O
can	O
learn	O
which	O
features	O
to	O
optimally	O
place	O
in	O
which	O
level	O
on	O
its	O
own	O
.	O
</s>
<s>
The	O
word	O
"	O
deep	O
"	O
in	O
"	O
deep	B-Algorithm
learning	I-Algorithm
"	O
refers	O
to	O
the	O
number	O
of	O
layers	O
through	O
which	O
the	O
data	O
is	O
transformed	O
.	O
</s>
<s>
More	O
precisely	O
,	O
deep	B-Algorithm
learning	I-Algorithm
systems	O
have	O
a	O
substantial	O
credit	O
assignment	O
path	O
(	O
CAP	O
)	O
depth	O
.	O
</s>
<s>
For	O
a	O
feedforward	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
,	O
the	O
depth	O
of	O
the	O
CAPs	O
is	O
that	O
of	O
the	O
network	O
and	O
is	O
the	O
number	O
of	O
hidden	O
layers	O
plus	O
one	O
(	O
as	O
the	O
output	O
layer	O
is	O
also	O
parameterized	O
)	O
.	O
</s>
<s>
For	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
in	O
which	O
a	O
signal	O
may	O
propagate	O
through	O
a	O
layer	O
more	O
than	O
once	O
,	O
the	O
CAP	O
depth	O
is	O
potentially	O
unlimited	O
.	O
</s>
<s>
No	O
universally	O
agreed-upon	O
threshold	O
of	O
depth	O
divides	O
shallow	O
learning	O
from	O
deep	B-Algorithm
learning	I-Algorithm
,	O
but	O
most	O
researchers	O
agree	O
that	O
deep	B-Algorithm
learning	I-Algorithm
involves	O
CAP	O
depth	O
higher	O
than	O
2	O
.	O
</s>
<s>
CAP	O
of	O
depth	O
2	O
has	O
been	O
shown	O
to	O
be	O
a	O
universal	B-Algorithm
approximator	I-Algorithm
in	O
the	O
sense	O
that	O
it	O
can	O
emulate	O
any	O
function	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
architectures	O
can	O
be	O
constructed	O
with	O
a	O
greedy	B-Algorithm
layer-by-layer	O
method	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
helps	O
to	O
disentangle	O
these	O
abstractions	O
and	O
pick	O
out	O
which	O
features	O
improve	O
performance	O
.	O
</s>
<s>
For	O
supervised	B-General_Concept
learning	I-General_Concept
tasks	O
,	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
eliminate	O
feature	B-General_Concept
engineering	I-General_Concept
,	O
by	O
translating	O
the	O
data	O
into	O
compact	O
intermediate	O
representations	O
akin	O
to	O
principal	B-Application
components	I-Application
,	O
and	O
derive	O
layered	O
structures	O
that	O
remove	O
redundancy	O
in	O
representation	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
algorithms	O
can	O
be	O
applied	O
to	O
unsupervised	B-General_Concept
learning	I-General_Concept
tasks	O
.	O
</s>
<s>
This	O
is	O
an	O
important	O
benefit	O
because	O
unlabeled	O
data	O
are	O
more	O
abundant	O
than	O
the	O
labeled	B-General_Concept
data	I-General_Concept
.	O
</s>
<s>
Examples	O
of	O
deep	O
structures	O
that	O
can	O
be	O
trained	O
in	O
an	O
unsupervised	B-General_Concept
manner	O
are	O
deep	B-Algorithm
belief	I-Algorithm
networks	I-Algorithm
.	O
</s>
<s>
Deep	O
neural	B-Architecture
networks	I-Architecture
are	O
generally	O
interpreted	O
in	O
terms	O
of	O
the	O
universal	B-Algorithm
approximation	I-Algorithm
theorem	I-Algorithm
or	O
probabilistic	O
inference	O
.	O
</s>
<s>
The	O
classic	O
universal	B-Algorithm
approximation	I-Algorithm
theorem	I-Algorithm
concerns	O
the	O
capacity	O
of	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
with	O
a	O
single	O
hidden	O
layer	O
of	O
finite	O
size	O
to	O
approximate	O
continuous	O
functions	O
.	O
</s>
<s>
In	O
1989	O
,	O
the	O
first	O
proof	O
was	O
published	O
by	O
George	O
Cybenko	O
for	O
sigmoid	B-Algorithm
activation	B-Algorithm
functions	I-Algorithm
and	O
was	O
generalised	O
to	O
feed-forward	O
multi-layer	O
architectures	O
in	O
1991	O
by	O
Kurt	O
Hornik	O
.	O
</s>
<s>
Recent	O
work	O
also	O
showed	O
that	O
universal	O
approximation	O
also	O
holds	O
for	O
non-bounded	O
activation	B-Algorithm
functions	I-Algorithm
such	O
as	O
Kunihiko	O
Fukushima	O
's	O
rectified	B-Algorithm
linear	I-Algorithm
unit	I-Algorithm
.	O
</s>
<s>
The	O
universal	B-Algorithm
approximation	I-Algorithm
theorem	I-Algorithm
for	O
deep	O
neural	B-Architecture
networks	I-Architecture
concerns	O
the	O
capacity	O
of	O
networks	O
with	O
bounded	O
width	O
but	O
the	O
depth	O
is	O
allowed	O
to	O
grow	O
.	O
</s>
<s>
proved	O
that	O
if	O
the	O
width	O
of	O
a	O
deep	O
neural	B-Architecture
network	I-Architecture
with	O
ReLU	B-Algorithm
activation	O
is	O
strictly	O
larger	O
than	O
the	O
input	O
dimension	O
,	O
then	O
the	O
network	O
can	O
approximate	O
any	O
Lebesgue	O
integrable	O
function	O
;	O
If	O
the	O
width	O
is	O
smaller	O
or	O
equal	O
to	O
the	O
input	O
dimension	O
,	O
then	O
a	O
deep	O
neural	B-Architecture
network	I-Architecture
is	O
not	O
a	O
universal	B-Algorithm
approximator	I-Algorithm
.	O
</s>
<s>
It	O
features	O
inference	O
,	O
as	O
well	O
as	O
the	O
optimization	O
concepts	O
of	O
training	O
and	O
testing	O
,	O
related	O
to	O
fitting	O
and	O
generalization	B-Algorithm
,	O
respectively	O
.	O
</s>
<s>
The	O
probabilistic	O
interpretation	O
led	O
to	O
the	O
introduction	O
of	O
dropout	B-Algorithm
as	O
regularizer	O
in	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
There	O
are	O
two	O
types	O
of	O
neural	B-Architecture
networks	I-Architecture
:	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
FNNs	O
)	O
and	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
RNNs	O
)	O
.	O
</s>
<s>
In	O
the	O
1920s	O
,	O
Wilhelm	O
Lenz	O
and	O
Ernst	O
Ising	O
created	O
and	O
analyzed	O
the	O
Ising	O
model	O
which	O
is	O
essentially	O
a	O
non-learning	O
RNN	B-Algorithm
architecture	O
consisting	O
of	O
neuron-like	O
threshold	O
elements	O
.	O
</s>
<s>
His	O
learning	O
RNN	B-Algorithm
was	O
popularised	O
by	O
John	O
Hopfield	O
in	O
1982	O
.	O
</s>
<s>
RNNs	O
have	O
become	O
central	O
for	O
speech	B-Application
recognition	I-Application
and	O
language	O
processing	O
.	O
</s>
<s>
Charles	O
Tappert	O
writes	O
that	O
Frank	O
Rosenblatt	O
developed	O
and	O
explored	O
all	O
of	O
the	O
basic	O
ingredients	O
of	O
the	O
deep	B-Algorithm
learning	I-Algorithm
systems	O
of	O
today	O
,	O
referring	O
to	O
Rosenblatt	O
's	O
1962	O
book	O
which	O
introduced	O
a	O
multilayer	B-Algorithm
perceptron	I-Algorithm
(	O
MLP	O
)	O
with	O
3	O
layers	O
:	O
an	O
input	O
layer	O
,	O
a	O
hidden	O
layer	O
with	O
randomized	O
weights	O
that	O
did	O
not	O
learn	O
,	O
and	O
an	O
output	O
layer	O
.	O
</s>
<s>
However	O
,	O
since	O
only	O
the	O
output	O
layer	O
had	O
learning	O
connections	O
,	O
this	O
was	O
not	O
yet	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
It	O
was	O
what	O
later	O
was	O
called	O
an	O
extreme	B-Algorithm
learning	I-Algorithm
machine	I-Algorithm
.	O
</s>
<s>
The	O
first	O
general	O
,	O
working	O
learning	O
algorithm	O
for	O
supervised	B-General_Concept
,	O
deep	O
,	O
feedforward	O
,	O
multilayer	B-Algorithm
perceptrons	I-Algorithm
was	O
published	O
by	O
Alexey	O
Ivakhnenko	O
and	O
Lapa	O
in	O
1967	O
.	O
</s>
<s>
A	O
1971	O
paper	O
described	O
a	O
deep	O
network	O
with	O
eight	O
layers	O
trained	O
by	O
the	O
group	B-Algorithm
method	I-Algorithm
of	I-Algorithm
data	I-Algorithm
handling	I-Algorithm
.	O
</s>
<s>
The	O
first	O
deep	B-Algorithm
learning	I-Algorithm
multilayer	B-Algorithm
perceptron	I-Algorithm
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>
In	O
1970	O
,	O
Seppo	O
Linnainmaa	O
published	O
the	O
reverse	O
mode	O
of	O
automatic	B-Algorithm
differentiation	I-Algorithm
of	O
discrete	O
connected	O
networks	O
of	O
nested	O
differentiable	O
functions	O
.	O
</s>
<s>
This	O
became	O
known	O
as	O
backpropagation	B-Algorithm
.	O
</s>
<s>
The	O
terminology	O
"	O
back-propagating	O
errors	O
"	O
was	O
actually	O
introduced	O
in	O
1962	O
by	O
Rosenblatt	O
,	O
but	O
he	O
did	O
not	O
know	O
how	O
to	O
implement	O
this	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>
Deep	B-Algorithm
learning	I-Algorithm
architectures	O
for	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
CNNs	B-Architecture
)	O
with	O
convolutional	O
layers	O
and	O
downsampling	O
layersbegan	O
with	O
the	O
Neocognitron	B-Algorithm
introduced	O
by	O
Kunihiko	O
Fukushima	O
in	O
1980	O
.	O
</s>
<s>
In	O
1969	O
,	O
he	O
also	O
introduced	O
the	O
ReLU	B-Algorithm
(	O
rectified	B-Algorithm
linear	I-Algorithm
unit	I-Algorithm
)	O
activation	B-Algorithm
function	I-Algorithm
.	O
</s>
<s>
The	O
rectifier	B-Algorithm
has	O
become	O
the	O
most	O
popular	O
activation	B-Algorithm
function	I-Algorithm
for	O
CNNs	B-Architecture
and	O
deep	B-Algorithm
learning	I-Algorithm
in	O
general	O
.	O
</s>
<s>
CNNs	B-Architecture
have	O
become	O
an	O
essential	O
tool	O
for	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
The	O
term	O
Deep	B-Algorithm
Learning	I-Algorithm
was	O
introduced	O
to	O
the	O
machine	O
learning	O
community	O
by	O
Rina	O
Dechter	O
in	O
1986	O
,	O
and	O
to	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
by	O
Igor	O
Aizenberg	O
and	O
colleagues	O
in	O
2000	O
,	O
in	O
the	O
context	O
of	O
Boolean	O
threshold	B-Algorithm
neurons	I-Algorithm
.	O
</s>
<s>
to	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
a	O
simplified	O
Neocognitron	B-Algorithm
with	O
convolutional	O
interconnections	O
between	O
the	O
image	O
feature	O
layers	O
and	O
the	O
last	O
fully	O
connected	O
layer	O
)	O
for	O
alphabet	O
recognition	O
.	O
</s>
<s>
They	O
also	O
proposed	O
an	O
implementation	O
of	O
the	O
CNN	B-Architecture
with	O
an	O
optical	O
computing	O
system	O
.	O
</s>
<s>
applied	O
backpropagation	B-Algorithm
to	O
a	O
CNN	B-Architecture
with	O
the	O
purpose	O
of	O
recognizing	O
handwritten	O
ZIP	O
codes	O
on	O
mail	O
.	O
</s>
<s>
LeNet-5	O
(	O
1998	O
)	O
,	O
a	O
7-level	O
CNN	B-Architecture
by	O
Yann	O
LeCun	O
et	O
al.	O
,	O
that	O
classifies	O
digits	O
,	O
was	O
applied	O
by	O
several	O
banks	O
to	O
recognize	O
hand-written	O
numbers	O
on	O
checks	O
digitized	O
in	O
32x32	O
pixel	O
images	O
.	O
</s>
<s>
In	O
the	O
1980s	O
,	O
backpropagation	B-Algorithm
did	O
not	O
work	O
well	O
for	O
deep	B-Algorithm
learning	I-Algorithm
with	O
long	O
credit	O
assignment	O
paths	O
.	O
</s>
<s>
To	O
overcome	O
this	O
problem	O
,	O
Juergen	O
Schmidhuber	O
(	O
1992	O
)	O
proposed	O
a	O
hierarchy	O
of	O
RNNs	O
pre-trained	O
one	O
level	O
at	O
a	O
time	O
by	O
self-supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
This	O
can	O
substantially	O
facilitate	O
downstream	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
The	O
RNN	B-Algorithm
hierarchy	O
can	O
be	O
collapsed	O
into	O
a	O
single	O
RNN	B-Algorithm
,	O
by	O
distilling	B-Algorithm
a	O
higher	O
level	O
chunker	O
network	O
into	O
a	O
lower	O
level	O
automatizer	O
network	O
.	O
</s>
<s>
In	O
1993	O
,	O
a	O
chunker	O
solved	O
a	O
deep	B-Algorithm
learning	I-Algorithm
task	O
whose	O
depth	O
exceeded	O
1000	O
.	O
</s>
<s>
In	O
1992	O
,	O
Juergen	O
Schmidhuber	O
also	O
published	O
an	O
alternative	O
to	O
RNNs	O
which	O
is	O
now	O
called	O
a	O
linear	O
Transformer	B-Algorithm
or	O
a	O
Transformer	B-Algorithm
with	O
linearized	O
self-attention	O
(	O
save	O
for	O
a	O
normalization	O
operator	O
)	O
.	O
</s>
<s>
It	O
learns	O
internal	O
spotlights	O
of	O
attention	O
:	O
a	O
slow	O
feedforward	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
learns	O
by	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
control	O
the	O
fast	O
weights	O
of	O
another	O
neural	B-Architecture
network	I-Architecture
through	O
outer	O
products	O
of	O
self-generated	O
activation	O
patterns	O
FROM	O
and	O
TO	O
(	O
which	O
are	O
now	O
called	O
key	O
and	O
value	O
for	O
self-attention	O
)	O
.	O
</s>
<s>
The	O
modern	O
Transformer	B-Algorithm
was	O
introduced	O
by	O
Ashish	O
Vaswani	O
et	O
.	O
</s>
<s>
It	O
combines	O
this	O
with	O
a	O
softmax	O
operator	O
and	O
a	O
projection	O
matrix	B-Architecture
.	O
</s>
<s>
Transformers	B-Algorithm
have	O
increasingly	O
become	O
the	O
model	O
of	O
choice	O
for	O
natural	B-Language
language	I-Language
processing	I-Language
.	O
</s>
<s>
Many	O
modern	O
large	O
language	B-Language
models	I-Language
such	O
as	O
ChatGPT	B-General_Concept
,	O
GPT-4	O
,	O
and	O
BERT	B-General_Concept
use	O
it	O
.	O
</s>
<s>
Transformers	B-Algorithm
are	O
also	O
increasingly	O
being	O
used	O
in	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
In	O
1991	O
,	O
Juergen	O
Schmidhuber	O
also	O
published	O
adversarial	O
neural	B-Architecture
networks	I-Architecture
that	O
contest	O
with	O
each	O
other	O
in	O
the	O
form	O
of	O
a	O
zero-sum	O
game	O
,	O
where	O
one	O
network	O
's	O
gain	O
is	O
the	O
other	O
network	O
's	O
loss	O
.	O
</s>
<s>
The	O
second	O
network	O
learns	O
by	O
gradient	B-Algorithm
descent	I-Algorithm
to	O
predict	O
the	O
reactions	O
of	O
the	O
environment	O
to	O
these	O
patterns	O
.	O
</s>
<s>
In	O
2014	O
,	O
this	O
principle	O
was	O
used	O
in	O
a	O
generative	B-Algorithm
adversarial	I-Algorithm
network	I-Algorithm
(	O
GAN	O
)	O
by	O
Ian	O
Goodfellow	O
et	O
al	O
.	O
</s>
<s>
This	O
can	O
be	O
used	O
to	O
create	O
realistic	O
deepfakes	B-Application
.	O
</s>
<s>
Excellent	O
image	O
quality	O
is	O
achieved	O
by	O
Nvidia	O
's	O
StyleGAN	B-Application
(	O
2018	O
)	O
based	O
on	O
the	O
Progressive	O
GAN	O
by	O
Tero	O
Karras	O
et	O
.	O
</s>
<s>
It	O
not	O
only	O
tested	O
the	O
neural	O
history	O
compressor	O
,	O
but	O
also	O
identified	O
and	O
analyzed	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
Hochreiter	O
proposed	O
recurrent	O
residual	B-Algorithm
connections	O
to	O
solve	O
this	O
problem	O
.	O
</s>
<s>
This	O
led	O
to	O
the	O
deep	B-Algorithm
learning	I-Algorithm
method	O
called	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
,	O
published	O
in	O
1997	O
.	O
</s>
<s>
LSTM	B-Algorithm
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
can	O
learn	O
"	O
very	O
deep	B-Algorithm
learning	I-Algorithm
"	O
tasks	O
with	O
long	O
credit	O
assignment	O
paths	O
that	O
require	O
memories	O
of	O
events	O
that	O
happened	O
thousands	O
of	O
discrete	O
time	O
steps	O
before	O
.	O
</s>
<s>
The	O
"	O
vanilla	O
LSTM	B-Algorithm
"	O
with	O
forget	O
gate	O
was	O
introduced	O
in	O
1999	O
by	O
Felix	O
Gers	O
,	O
Schmidhuber	O
and	O
Fred	O
Cummins	O
.	O
</s>
<s>
LSTM	B-Algorithm
has	O
become	O
the	O
most	O
cited	O
neural	B-Architecture
network	I-Architecture
of	O
the	O
20th	O
century	O
.	O
</s>
<s>
In	O
2015	O
,	O
Rupesh	O
Kumar	O
Srivastava	O
,	O
Klaus	O
Greff	O
,	O
and	O
Schmidhuber	O
used	O
LSTM	B-Algorithm
principles	O
to	O
create	O
the	O
Highway	B-General_Concept
network	I-General_Concept
,	O
a	O
feedforward	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
with	O
hundreds	O
of	O
layers	O
,	O
much	O
deeper	O
than	O
previous	O
networks	O
.	O
</s>
<s>
7	O
months	O
later	O
,	O
Kaiming	O
He	O
,	O
Xiangyu	O
Zhang	O
;	O
Shaoqing	O
Ren	O
,	O
and	O
Jian	O
Sun	O
won	O
the	O
ImageNet	O
2015	O
competition	O
with	O
an	O
open-gated	O
or	O
gateless	O
Highway	B-General_Concept
network	I-General_Concept
variant	O
called	O
Residual	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
.	O
</s>
<s>
This	O
has	O
become	O
the	O
most	O
cited	O
neural	B-Architecture
network	I-Architecture
of	O
the	O
21st	O
century	O
.	O
</s>
<s>
In	O
1994	O
,	O
André	O
de	O
Carvalho	O
,	O
together	O
with	O
Mike	O
Fairhurst	O
and	O
David	O
Bisset	O
,	O
published	O
experimental	O
results	O
of	O
a	O
multi-layer	O
boolean	O
neural	B-Architecture
network	I-Architecture
,	O
also	O
known	O
as	O
a	O
weightless	O
neural	B-Architecture
network	I-Architecture
,	O
composed	O
of	O
a	O
3-layers	O
self-organising	O
feature	O
extraction	O
neural	B-Architecture
network	I-Architecture
module	O
(	O
SOFT	O
)	O
followed	O
by	O
a	O
multi-layer	O
classification	O
neural	B-Architecture
network	I-Architecture
module	O
(	O
GSN	O
)	O
,	O
which	O
were	O
independently	O
trained	O
.	O
</s>
<s>
In	O
1995	O
,	O
Brendan	O
Frey	O
demonstrated	O
that	O
it	O
was	O
possible	O
to	O
train	O
(	O
over	O
two	O
days	O
)	O
a	O
network	O
containing	O
six	O
fully	O
connected	O
layers	O
and	O
several	O
hundred	O
hidden	O
units	O
using	O
the	O
wake-sleep	B-Algorithm
algorithm	I-Algorithm
,	O
co-developed	O
with	O
Peter	O
Dayan	O
and	O
Hinton	O
.	O
</s>
<s>
Simpler	O
models	O
that	O
use	O
task-specific	O
handcrafted	O
features	O
such	O
as	O
Gabor	O
filters	O
and	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
(	O
SVMs	B-Algorithm
)	O
were	O
a	O
popular	O
choice	O
in	O
the	O
1990s	O
and	O
2000s	O
,	O
because	O
of	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
's	O
(	O
ANN	O
)	O
computational	O
cost	O
and	O
a	O
lack	O
of	O
understanding	O
of	O
how	O
the	O
brain	O
wires	O
its	O
biological	O
networks	O
.	O
</s>
<s>
Both	O
shallow	O
and	O
deep	B-Algorithm
learning	I-Algorithm
(	O
e.g.	O
,	O
recurrent	O
nets	O
)	O
of	O
ANNs	O
for	O
speech	B-Application
recognition	I-Application
have	O
been	O
explored	O
for	O
many	O
years	O
.	O
</s>
<s>
Most	O
speech	B-Application
recognition	I-Application
researchers	O
moved	O
away	O
from	O
neural	B-Architecture
nets	I-Architecture
to	O
pursue	O
generative	O
modeling	O
.	O
</s>
<s>
Funded	O
by	O
the	O
US	O
government	O
's	O
NSA	O
and	O
DARPA	O
,	O
SRI	O
studied	O
deep	O
neural	B-Architecture
networks	I-Architecture
in	O
speech	O
and	O
speaker	B-Application
recognition	I-Application
.	O
</s>
<s>
The	O
speaker	B-Application
recognition	I-Application
team	O
led	O
by	O
Larry	O
Heck	O
reported	O
significant	O
success	O
with	O
deep	O
neural	B-Architecture
networks	I-Architecture
in	O
speech	O
processing	O
in	O
the	O
1998	O
National	O
Institute	O
of	O
Standards	O
and	O
Technology	O
Speaker	B-Application
Recognition	I-Application
evaluation	O
.	O
</s>
<s>
The	O
SRI	O
deep	O
neural	B-Architecture
network	I-Architecture
was	O
then	O
deployed	O
in	O
the	O
Nuance	O
Verifier	O
,	O
representing	O
the	O
first	O
major	O
industrial	O
application	O
of	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
The	O
principle	O
of	O
elevating	O
"	O
raw	O
"	O
features	O
over	O
hand-crafted	O
optimization	O
was	O
first	O
explored	O
successfully	O
in	O
the	O
architecture	O
of	O
deep	O
autoencoder	B-Algorithm
on	O
the	O
"	O
raw	O
"	O
spectrogram	O
or	O
linear	O
filter-bank	O
features	O
in	O
the	O
late	O
1990s	O
,	O
showing	O
its	O
superiority	O
over	O
the	O
Mel-Cepstral	O
features	O
that	O
contain	O
stages	O
of	O
fixed	O
transformation	O
from	O
spectrograms	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
was	O
taken	O
over	O
by	O
LSTM	B-Algorithm
.	O
</s>
<s>
In	O
2003	O
,	O
LSTM	B-Algorithm
started	O
to	O
become	O
competitive	O
with	O
traditional	O
speech	B-Application
recognizers	I-Application
on	O
certain	O
tasks	O
.	O
</s>
<s>
In	O
2006	O
,	O
Alex	O
Graves	O
,	O
Santiago	O
Fernández	O
,	O
Faustino	O
Gomez	O
,	O
and	O
Schmidhuber	O
combined	O
it	O
with	O
connectionist	B-Algorithm
temporal	I-Algorithm
classification	I-Algorithm
(	O
CTC	O
)	O
in	O
stacks	O
of	O
LSTM	B-Algorithm
RNNs	O
.	O
</s>
<s>
In	O
2015	O
,	O
Google	B-Application
's	I-Application
speech	B-Application
recognition	I-Application
reportedly	O
experienced	O
a	O
dramatic	O
performance	O
jump	O
of	O
49%	O
through	O
CTC-trained	O
LSTM	B-Algorithm
,	O
which	O
they	O
made	O
available	O
through	O
Google	B-Application
Voice	I-Application
Search	I-Application
.	O
</s>
<s>
The	O
impact	O
of	O
deep	B-Algorithm
learning	I-Algorithm
in	O
industry	O
began	O
in	O
the	O
early	O
2000s	O
,	O
when	O
CNNs	B-Architecture
already	O
processed	O
an	O
estimated	O
10%	O
to	O
20%	O
of	O
all	O
the	O
checks	O
written	O
in	O
the	O
US	O
,	O
according	O
to	O
Yann	O
LeCun	O
.	O
</s>
<s>
Industrial	O
applications	O
of	O
deep	B-Algorithm
learning	I-Algorithm
to	O
large-scale	O
speech	B-Application
recognition	I-Application
started	O
around	O
2010	O
.	O
</s>
<s>
In	O
2006	O
,	O
publications	O
by	O
Geoff	O
Hinton	O
,	O
Ruslan	O
Salakhutdinov	O
,	O
Osindero	O
and	O
Teh	O
showed	O
how	O
a	O
many-layered	O
feedforward	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
could	O
be	O
effectively	O
pre-trained	O
one	O
layer	O
at	O
a	O
time	O
,	O
treating	O
each	O
layer	O
in	O
turn	O
as	O
an	O
unsupervised	B-General_Concept
restricted	B-Algorithm
Boltzmann	I-Algorithm
machine	I-Algorithm
,	O
then	O
fine-tuning	O
it	O
using	O
supervised	B-General_Concept
backpropagation	B-Algorithm
.	O
</s>
<s>
The	O
papers	O
referred	O
to	O
learning	O
for	O
deep	B-Algorithm
belief	I-Algorithm
nets	I-Algorithm
.	O
</s>
<s>
The	O
2009	O
NIPS	O
Workshop	O
on	O
Deep	B-Algorithm
Learning	I-Algorithm
for	O
Speech	B-Application
Recognition	I-Application
was	O
motivated	O
by	O
the	O
limitations	O
of	O
deep	O
generative	O
models	O
of	O
speech	O
,	O
and	O
the	O
possibility	O
that	O
given	O
more	O
capable	O
hardware	B-Architecture
and	O
large-scale	O
data	O
sets	O
that	O
deep	O
neural	B-Architecture
nets	I-Architecture
(	O
DNN	O
)	O
might	O
become	O
practical	O
.	O
</s>
<s>
It	O
was	O
believed	O
that	O
pre-training	O
DNNs	O
using	O
generative	O
models	O
of	O
deep	B-Algorithm
belief	I-Algorithm
nets	I-Algorithm
(	O
DBN	O
)	O
would	O
overcome	O
the	O
main	O
difficulties	O
of	O
neural	B-Architecture
nets	I-Architecture
.	O
</s>
<s>
However	O
,	O
it	O
was	O
discovered	O
that	O
replacing	O
pre-training	O
with	O
large	O
amounts	O
of	O
training	O
data	O
for	O
straightforward	O
backpropagation	B-Algorithm
when	O
using	O
DNNs	O
with	O
large	O
,	O
context-dependent	O
output	O
layers	O
produced	O
error	O
rates	O
dramatically	O
lower	O
than	O
then-state-of-the-art	O
Gaussian	O
mixture	O
model	O
(	O
GMM	O
)	O
/Hidden	O
Markov	O
Model	O
(	O
HMM	O
)	O
and	O
also	O
than	O
more-advanced	O
generative	O
model-based	O
systems	O
.	O
</s>
<s>
The	O
nature	O
of	O
the	O
recognition	O
errors	O
produced	O
by	O
the	O
two	O
types	O
of	O
systems	O
was	O
characteristically	O
different	O
,	O
offering	O
technical	O
insights	O
into	O
how	O
to	O
integrate	O
deep	B-Algorithm
learning	I-Algorithm
into	O
the	O
existing	O
highly	O
efficient	O
,	O
run-time	O
speech	O
decoding	O
system	O
deployed	O
by	O
all	O
major	O
speech	B-Application
recognition	I-Application
systems	O
.	O
</s>
<s>
Analysis	O
around	O
2009	O
–	O
2010	O
,	O
contrasting	O
the	O
GMM	O
(	O
and	O
other	O
generative	O
speech	O
models	O
)	O
vs.	O
DNN	O
models	O
,	O
stimulated	O
early	O
industrial	O
investment	O
in	O
deep	B-Algorithm
learning	I-Algorithm
for	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
In	O
2010	O
,	O
researchers	O
extended	O
deep	B-Algorithm
learning	I-Algorithm
from	O
TIMIT	B-Application
to	O
large	O
vocabulary	O
speech	B-Application
recognition	I-Application
,	O
by	O
adopting	O
large	O
output	O
layers	O
of	O
the	O
DNN	O
based	O
on	O
context-dependent	O
HMM	O
states	O
constructed	O
by	O
decision	B-Algorithm
trees	I-Algorithm
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
part	O
of	O
state-of-the-art	O
systems	O
in	O
various	O
disciplines	O
,	O
particularly	O
computer	B-Application
vision	I-Application
and	O
automatic	B-Application
speech	I-Application
recognition	I-Application
(	O
ASR	O
)	O
.	O
</s>
<s>
Results	O
on	O
commonly	O
used	O
evaluation	O
sets	O
such	O
as	O
TIMIT	B-Application
(	O
ASR	O
)	O
and	O
MNIST	B-General_Concept
(	O
image	O
classification	O
)	O
,	O
as	O
well	O
as	O
a	O
range	O
of	O
large-vocabulary	O
speech	B-Application
recognition	I-Application
tasks	O
have	O
steadily	O
improved	O
.	O
</s>
<s>
Convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
CNNs	B-Architecture
)	O
were	O
superseded	O
for	O
ASR	O
by	O
CTC	O
for	O
LSTM	B-Algorithm
.	O
</s>
<s>
but	O
are	O
more	O
successful	O
in	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
Advances	O
in	O
hardware	B-Architecture
have	O
driven	O
renewed	O
interest	O
in	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
In	O
2009	O
,	O
Nvidia	O
was	O
involved	O
in	O
what	O
was	O
called	O
the	O
“	O
big	O
bang	O
”	O
of	O
deep	B-Algorithm
learning	I-Algorithm
,	O
“	O
as	O
deep-learning	B-Algorithm
neural	B-Architecture
networks	I-Architecture
were	O
trained	O
with	O
Nvidia	O
graphics	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
.	O
”	O
That	O
year	O
,	O
Andrew	O
Ng	O
determined	O
that	O
GPUs	B-Architecture
could	O
increase	O
the	O
speed	O
of	O
deep-learning	B-Algorithm
systems	O
by	O
about	O
100	O
times	O
.	O
</s>
<s>
In	O
particular	O
,	O
GPUs	B-Architecture
are	O
well-suited	O
for	O
the	O
matrix/vector	O
computations	O
involved	O
in	O
machine	O
learning	O
.	O
</s>
<s>
GPUs	B-Architecture
speed	O
up	O
training	O
algorithms	O
by	O
orders	O
of	O
magnitude	O
,	O
reducing	O
running	O
times	O
from	O
weeks	O
to	O
days	O
.	O
</s>
<s>
Further	O
,	O
specialized	O
hardware	B-Architecture
and	O
algorithm	O
optimizations	O
can	O
be	O
used	O
for	O
efficient	O
processing	O
of	O
deep	B-Algorithm
learning	I-Algorithm
models	O
.	O
</s>
<s>
In	O
the	O
late	O
2000s	O
,	O
deep	B-Algorithm
learning	I-Algorithm
started	O
to	O
outperform	O
other	O
methods	O
in	O
machine	O
learning	O
competitions	O
.	O
</s>
<s>
In	O
2009	O
,	O
a	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
trained	O
by	O
connectionist	B-Algorithm
temporal	I-Algorithm
classification	I-Algorithm
(	O
Alex	O
Graves	O
,	O
Santiago	O
Fernández	O
,	O
Faustino	O
Gomez	O
,	O
and	O
Juergen	O
Schmidhuber	O
,	O
2006	O
)	O
was	O
the	O
first	O
RNN	B-Algorithm
to	O
win	O
pattern	O
recognition	O
contests	O
,	O
winning	O
three	O
competitions	O
in	O
connected	O
handwriting	B-Application
recognition	I-Application
.	O
</s>
<s>
Google	B-Application
later	O
used	O
CTC-trained	O
LSTM	B-Algorithm
for	O
speech	B-Application
recognition	I-Application
on	O
the	O
smartphone	B-Application
.	O
</s>
<s>
Although	O
CNNs	B-Architecture
trained	O
by	O
backpropagation	B-Algorithm
had	O
been	O
around	O
for	O
decades	O
,	O
and	O
GPU	B-Architecture
implementations	O
of	O
NNs	O
for	O
years	O
,	O
including	O
CNNs	B-Architecture
,	O
faster	O
implementations	O
of	O
CNNs	B-Architecture
on	O
GPUs	B-Architecture
were	O
needed	O
to	O
progress	O
on	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
Until	O
2011	O
,	O
CNNs	B-Architecture
did	O
not	O
play	O
a	O
major	O
role	O
at	O
computer	B-Application
vision	I-Application
conferences	O
,	O
but	O
in	O
June	O
2012	O
,	O
a	O
paper	O
by	O
Ciresan	O
et	O
al	O
.	O
</s>
<s>
at	O
the	O
leading	O
conference	O
CVPR	O
showed	O
how	O
max-pooling	O
CNNs	B-Architecture
on	O
GPU	B-Architecture
can	O
dramatically	O
improve	O
many	O
vision	O
benchmark	O
records	O
.	O
</s>
<s>
In	O
October	O
2012	O
,	O
the	O
similar	O
AlexNet	B-Algorithm
by	O
Alex	O
Krizhevsky	O
,	O
Ilya	O
Sutskever	O
,	O
and	O
Geoffrey	O
Hinton	O
won	O
the	O
large-scale	O
ImageNet	O
competition	O
by	O
a	O
significant	O
margin	O
over	O
shallow	O
machine	O
learning	O
methods	O
.	O
</s>
<s>
won	O
the	O
ImageNet	O
2014	O
competition	O
,	O
following	O
a	O
similar	O
trend	O
in	O
large-scale	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
Image	O
classification	O
was	O
then	O
extended	O
to	O
the	O
more	O
challenging	O
task	O
of	O
generating	B-Application
descriptions	I-Application
(	O
captions	O
)	O
for	O
images	O
,	O
often	O
as	O
a	O
combination	O
of	O
CNNs	B-Architecture
and	O
LSTMs	B-Algorithm
.	O
</s>
<s>
In	O
2012	O
,	O
a	O
team	O
led	O
by	O
George	O
E	O
.	O
Dahl	O
won	O
the	O
"	O
Merck	O
Molecular	O
Activity	O
Challenge	O
"	O
using	O
multi-task	B-General_Concept
deep	O
neural	B-Architecture
networks	I-Architecture
to	O
predict	O
the	O
biomolecular	O
target	O
of	O
one	O
drug	O
.	O
</s>
<s>
In	O
2014	O
,	O
Sepp	O
Hochreiter	O
's	O
group	O
used	O
deep	B-Algorithm
learning	I-Algorithm
to	O
detect	O
off-target	O
and	O
toxic	O
effects	O
of	O
environmental	O
chemicals	O
in	O
nutrients	O
,	O
household	O
products	O
and	O
drugs	O
and	O
won	O
the	O
"	O
Tox21	O
Data	O
Challenge	O
"	O
of	O
NIH	O
,	O
FDA	O
and	O
NCATS	O
.	O
</s>
<s>
In	O
2016	O
,	O
Roger	O
Parloff	O
mentioned	O
a	O
"	O
deep	B-Algorithm
learning	I-Algorithm
revolution	O
"	O
that	O
has	O
transformed	O
the	O
AI	B-Application
industry	O
.	O
</s>
<s>
In	O
March	O
2019	O
,	O
Yoshua	O
Bengio	O
,	O
Geoffrey	O
Hinton	O
and	O
Yann	O
LeCun	O
were	O
awarded	O
the	O
Turing	O
Award	O
for	O
conceptual	O
and	O
engineering	O
breakthroughs	O
that	O
have	O
made	O
deep	O
neural	B-Architecture
networks	I-Architecture
a	O
critical	O
component	O
of	O
computing	O
.	O
</s>
<s>
Artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
ANNs	O
)	O
or	O
connectionist	O
systems	O
are	O
computing	O
systems	O
inspired	O
by	O
the	O
biological	B-General_Concept
neural	I-General_Concept
networks	I-General_Concept
that	O
constitute	O
animal	O
brains	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
image	O
recognition	O
,	O
they	O
might	O
learn	O
to	O
identify	O
images	O
that	O
contain	O
cats	O
by	O
analyzing	O
example	O
images	O
that	O
have	O
been	O
manually	O
labeled	B-General_Concept
as	O
"	O
cat	O
"	O
or	O
"	O
no	O
cat	O
"	O
and	O
using	O
the	O
analytic	O
results	O
to	O
identify	O
cats	O
in	O
other	O
images	O
.	O
</s>
<s>
They	O
have	O
found	O
most	O
use	O
in	O
applications	O
difficult	O
to	O
express	O
with	O
a	O
traditional	O
computer	O
algorithm	O
using	O
rule-based	B-Language
programming	I-Language
.	O
</s>
<s>
An	O
ANN	O
is	O
based	O
on	O
a	O
collection	O
of	O
connected	O
units	O
called	O
artificial	B-Algorithm
neurons	I-Algorithm
,	O
(	O
analogous	O
to	O
biological	O
neurons	O
in	O
a	O
biological	O
brain	O
)	O
.	O
</s>
<s>
Each	O
connection	O
(	O
synapse	B-Application
)	O
between	O
neurons	O
can	O
transmit	O
a	O
signal	O
to	O
another	O
neuron	O
.	O
</s>
<s>
Neurons	O
and	O
synapses	B-Application
may	O
also	O
have	O
a	O
weight	O
that	O
varies	O
as	O
learning	O
proceeds	O
,	O
which	O
can	O
increase	O
or	O
decrease	O
the	O
strength	O
of	O
the	O
signal	O
that	O
it	O
sends	O
downstream	O
.	O
</s>
<s>
The	O
original	O
goal	O
of	O
the	O
neural	B-Architecture
network	I-Architecture
approach	O
was	O
to	O
solve	O
problems	O
in	O
the	O
same	O
way	O
that	O
a	O
human	O
brain	O
would	O
.	O
</s>
<s>
Over	O
time	O
,	O
attention	O
focused	O
on	O
matching	O
specific	O
mental	O
abilities	O
,	O
leading	O
to	O
deviations	O
from	O
biology	O
such	O
as	O
backpropagation	B-Algorithm
,	O
or	O
passing	O
information	O
in	O
the	O
reverse	O
direction	O
and	O
adjusting	O
the	O
network	O
to	O
reflect	O
that	O
information	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
have	O
been	O
used	O
on	O
a	O
variety	O
of	O
tasks	O
,	O
including	O
computer	B-Application
vision	I-Application
,	O
speech	B-Application
recognition	I-Application
,	O
machine	B-Application
translation	I-Application
,	O
social	O
network	O
filtering	O
,	O
playing	B-Algorithm
board	I-Algorithm
and	I-Algorithm
video	I-Algorithm
games	I-Algorithm
and	O
medical	O
diagnosis	O
.	O
</s>
<s>
As	O
of	O
2017	O
,	O
neural	B-Architecture
networks	I-Architecture
typically	O
have	O
a	O
few	O
thousand	O
to	O
a	O
few	O
million	O
units	O
and	O
millions	O
of	O
connections	O
.	O
</s>
<s>
A	O
deep	O
neural	B-Architecture
network	I-Architecture
(	O
DNN	O
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
ANN	O
)	O
with	O
multiple	O
layers	O
between	O
the	O
input	O
and	O
output	O
layers	O
.	O
</s>
<s>
There	O
are	O
different	O
types	O
of	O
neural	B-Architecture
networks	I-Architecture
but	O
they	O
always	O
consist	O
of	O
the	O
same	O
components	O
:	O
neurons	O
,	O
synapses	B-Application
,	O
weights	O
,	O
biases	O
,	O
and	O
functions	O
.	O
</s>
<s>
Recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
RNNs	O
)	O
,	O
in	O
which	O
data	O
can	O
flow	O
in	O
any	O
direction	O
,	O
are	O
used	O
for	O
applications	O
such	O
as	O
language	B-Language
modeling	I-Language
.	O
</s>
<s>
Long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
is	O
particularly	O
effective	O
for	O
this	O
use	O
.	O
</s>
<s>
Convolutional	O
deep	O
neural	B-Architecture
networks	I-Architecture
(	O
CNNs	B-Architecture
)	O
are	O
used	O
in	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
CNNs	B-Architecture
also	O
have	O
been	O
applied	O
to	O
acoustic	O
modeling	O
for	O
automatic	B-Application
speech	I-Application
recognition	I-Application
(	O
ASR	O
)	O
.	O
</s>
<s>
Two	O
common	O
issues	O
are	O
overfitting	B-Error_Name
and	O
computation	O
time	O
.	O
</s>
<s>
DNNs	O
are	O
prone	O
to	O
overfitting	B-Error_Name
because	O
of	O
the	O
added	O
layers	O
of	O
abstraction	O
,	O
which	O
allow	O
them	O
to	O
model	O
rare	O
dependencies	O
in	O
the	O
training	O
data	O
.	O
</s>
<s>
Regularization	O
methods	O
such	O
as	O
Ivakhnenko	O
's	O
unit	O
pruning	O
or	O
weight	O
decay	O
(	O
-regularization	O
)	O
or	O
sparsity	B-Algorithm
(	O
-regularization	O
)	O
can	O
be	O
applied	O
during	O
training	O
to	O
combat	O
overfitting	B-Error_Name
.	O
</s>
<s>
Alternatively	O
dropout	B-Algorithm
regularization	O
randomly	O
omits	O
units	O
from	O
the	O
hidden	O
layers	O
during	O
training	O
.	O
</s>
<s>
Finally	O
,	O
data	O
can	O
be	O
augmented	O
via	O
methods	O
such	O
as	O
cropping	O
and	O
rotating	O
such	O
that	O
smaller	O
training	O
sets	O
can	O
be	O
increased	O
in	O
size	O
to	O
reduce	O
the	O
chances	O
of	O
overfitting	B-Error_Name
.	O
</s>
<s>
DNNs	O
must	O
consider	O
many	O
training	O
parameters	O
,	O
such	O
as	O
the	O
size	O
(	O
number	O
of	O
layers	O
and	O
number	O
of	O
units	O
per	O
layer	O
)	O
,	O
the	O
learning	B-General_Concept
rate	I-General_Concept
,	O
and	O
initial	O
weights	O
.	O
</s>
<s>
Large	O
processing	O
capabilities	O
of	O
many-core	O
architectures	O
(	O
such	O
as	O
GPUs	B-Architecture
or	O
the	O
Intel	O
Xeon	O
Phi	O
)	O
have	O
produced	O
significant	O
speedups	O
in	O
training	O
,	O
because	O
of	O
the	O
suitability	O
of	O
such	O
processing	O
architectures	O
for	O
the	O
matrix	B-Architecture
and	O
vector	O
computations	O
.	O
</s>
<s>
Alternatively	O
,	O
engineers	O
may	O
look	O
for	O
other	O
types	O
of	O
neural	B-Architecture
networks	I-Architecture
with	O
more	O
straightforward	O
and	O
convergent	O
training	O
algorithms	O
.	O
</s>
<s>
CMAC	O
(	O
cerebellar	B-Algorithm
model	I-Algorithm
articulation	I-Algorithm
controller	I-Algorithm
)	O
is	O
one	O
such	O
kind	O
of	O
neural	B-Architecture
network	I-Architecture
.	O
</s>
<s>
It	O
does	O
n't	O
require	O
learning	B-General_Concept
rates	I-General_Concept
or	O
randomized	O
initial	O
weights	O
for	O
CMAC	O
.	O
</s>
<s>
Since	O
the	O
2010s	O
,	O
advances	O
in	O
both	O
machine	O
learning	O
algorithms	O
and	O
computer	B-Architecture
hardware	I-Architecture
have	O
led	O
to	O
more	O
efficient	O
methods	O
for	O
training	O
deep	O
neural	B-Architecture
networks	I-Architecture
that	O
contain	O
many	O
layers	O
of	O
non-linear	O
hidden	O
units	O
and	O
a	O
very	O
large	O
output	O
layer	O
.	O
</s>
<s>
By	O
2019	O
,	O
graphic	B-Architecture
processing	I-Architecture
units	I-Architecture
(	O
GPUs	B-Architecture
)	O
,	O
often	O
with	O
AI-specific	O
enhancements	O
,	O
had	O
displaced	O
CPUs	O
as	O
the	O
dominant	O
method	O
of	O
training	O
large-scale	O
commercial	O
cloud	O
AI	B-Application
.	O
</s>
<s>
OpenAI	O
estimated	O
the	O
hardware	B-Architecture
computation	O
used	O
in	O
the	O
largest	O
deep	B-Algorithm
learning	I-Algorithm
projects	O
from	O
AlexNet	B-Algorithm
(	O
2012	O
)	O
to	O
AlphaZero	O
(	O
2017	O
)	O
,	O
and	O
found	O
a	O
300,000	O
-fold	O
increase	O
in	O
the	O
amount	O
of	O
computation	O
required	O
,	O
with	O
a	O
doubling-time	O
trendline	O
of	O
3.4	O
months	O
.	O
</s>
<s>
Special	O
electronic	O
circuits	O
called	O
deep	B-Algorithm
learning	I-Algorithm
processors	I-Algorithm
were	O
designed	O
to	O
speed	O
up	O
deep	B-Algorithm
learning	I-Algorithm
algorithms	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
processors	I-Algorithm
include	O
neural	O
processing	O
units	O
(	O
NPUs	O
)	O
in	O
Huawei	O
cellphones	O
and	O
cloud	B-Architecture
computing	I-Architecture
servers	O
such	O
as	O
tensor	B-Device
processing	I-Device
units	I-Device
(	O
TPU	O
)	O
in	O
the	B-Application
Google	I-Application
Cloud	O
Platform	O
.	O
</s>
<s>
Cerebras	O
Systems	O
has	O
also	O
built	O
a	O
dedicated	O
system	O
to	O
handle	O
large	O
deep	B-Algorithm
learning	I-Algorithm
models	O
,	O
the	O
CS-2	O
,	O
based	O
on	O
the	O
largest	O
processor	O
in	O
the	O
industry	O
,	O
the	O
second-generation	O
Wafer	O
Scale	O
Engine	O
(	O
WSE-2	O
)	O
.	O
</s>
<s>
Atomically	O
thin	O
semiconductors	O
are	O
considered	O
promising	O
for	O
energy-efficient	O
deep	B-Algorithm
learning	I-Algorithm
hardware	B-Architecture
where	O
the	O
same	O
basic	O
device	O
structure	O
is	O
used	O
for	O
both	O
logic	O
operations	O
and	O
data	O
storage	O
.	O
</s>
<s>
published	O
experiments	O
with	O
a	O
large-area	O
active	O
channel	O
material	O
for	O
developing	O
logic-in-memory	O
devices	O
and	O
circuits	O
based	O
on	O
floating-gate	B-Algorithm
field-effect	O
transistors	O
(	O
FGFETs	O
)	O
.	O
</s>
<s>
proposed	O
an	O
integrated	O
photonic	O
hardware	B-General_Concept
accelerator	I-General_Concept
for	O
parallel	O
convolutional	O
processing	O
.	O
</s>
<s>
The	O
authors	O
identify	O
two	O
key	O
advantages	O
of	O
integrated	O
photonics	O
over	O
its	O
electronic	O
counterparts	O
:	O
(	O
1	O
)	O
massively	O
parallel	O
data	O
transfer	O
through	O
wavelength	O
division	O
multiplexing	B-Architecture
in	O
conjunction	O
with	O
frequency	O
combs	O
,	O
and	O
(	O
2	O
)	O
extremely	O
high	O
data	O
modulation	O
speeds	O
.	O
</s>
<s>
Their	O
system	O
can	O
execute	O
trillions	O
of	O
multiply-accumulate	O
operations	O
per	O
second	O
,	O
indicating	O
the	O
potential	O
of	O
integrated	O
photonics	O
in	O
data-heavy	O
AI	B-Application
applications	I-Application
.	O
</s>
<s>
Large-scale	O
automatic	B-Application
speech	I-Application
recognition	I-Application
is	O
the	O
first	O
and	O
most	O
convincing	O
successful	O
case	O
of	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
LSTM	B-Algorithm
RNNs	O
can	O
learn	O
"	O
Very	O
Deep	B-Algorithm
Learning	I-Algorithm
"	O
tasks	O
that	O
involve	O
multi-second	O
intervals	O
containing	O
speech	O
events	O
separated	O
by	O
thousands	O
of	O
discrete	O
time	O
steps	O
,	O
where	O
one	O
time	O
step	O
corresponds	O
to	O
about	O
10	O
ms.	O
LSTM	B-Algorithm
with	O
forget	O
gates	O
is	O
competitive	O
with	O
traditional	O
speech	B-Application
recognizers	I-Application
on	O
certain	O
tasks	O
.	O
</s>
<s>
The	O
initial	O
success	O
in	O
speech	B-Application
recognition	I-Application
was	O
based	O
on	O
small-scale	O
recognition	O
tasks	O
based	O
on	O
TIMIT	B-Application
.	O
</s>
<s>
More	O
importantly	O
,	O
the	O
TIMIT	B-Application
task	O
concerns	O
phone-sequence	O
recognition	O
,	O
which	O
,	O
unlike	O
word-sequence	O
recognition	O
,	O
allows	O
weak	O
phone	O
bigram	O
language	B-Language
models	I-Language
.	O
</s>
<s>
This	O
lets	O
the	O
strength	O
of	O
the	O
acoustic	O
modeling	O
aspects	O
of	O
speech	B-Application
recognition	I-Application
be	O
more	O
easily	O
analyzed	O
.	O
</s>
<s>
The	O
debut	O
of	O
DNNs	O
for	O
speaker	B-Application
recognition	I-Application
in	O
the	O
late	O
1990s	O
and	O
speech	B-Application
recognition	I-Application
around	O
2009-2011	O
and	O
of	O
LSTM	B-Algorithm
around	O
2003	O
–	O
2007	O
,	O
accelerated	O
progress	O
in	O
eight	O
major	O
areas	O
:	O
</s>
<s>
All	O
major	O
commercial	O
speech	B-Application
recognition	I-Application
systems	O
(	O
e.g.	O
,	O
Microsoft	B-Application
Cortana	I-Application
,	O
Xbox	B-Protocol
,	O
Skype	B-Protocol
Translator	I-Protocol
,	O
Amazon	B-Application
Alexa	I-Application
,	O
Google	B-Application
Now	I-Application
,	O
Apple	B-Application
Siri	I-Application
,	O
Baidu	B-Application
and	O
iFlyTek	O
voice	B-Application
search	I-Application
,	O
and	O
a	O
range	O
of	O
Nuance	O
speech	O
products	O
,	O
etc	O
.	O
)	O
</s>
<s>
are	O
based	O
on	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
A	O
common	O
evaluation	O
set	O
for	O
image	O
classification	O
is	O
the	O
MNIST	B-General_Concept
database	I-General_Concept
data	O
set	O
.	O
</s>
<s>
MNIST	B-General_Concept
is	O
composed	O
of	O
handwritten	O
digits	O
and	O
includes	O
60,000	O
training	O
examples	O
and	O
10,000	O
test	O
examples	O
.	O
</s>
<s>
As	O
with	O
TIMIT	B-Application
,	O
its	O
small	O
size	O
lets	O
users	O
test	O
multiple	O
configurations	O
.	O
</s>
<s>
Closely	O
related	O
to	O
the	O
progress	O
that	O
has	O
been	O
made	O
in	O
image	O
recognition	O
is	O
the	O
increasing	O
application	O
of	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
to	O
various	O
visual	O
art	O
tasks	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
have	O
been	O
used	O
for	O
implementing	O
language	B-Language
models	I-Language
since	O
the	O
early	O
2000s	O
.	O
</s>
<s>
LSTM	B-Algorithm
helped	O
to	O
improve	O
machine	B-Application
translation	I-Application
and	O
language	B-Language
modeling	I-Language
.	O
</s>
<s>
Other	O
key	O
techniques	O
in	O
this	O
field	O
are	O
negative	O
sampling	O
and	O
word	B-General_Concept
embedding	I-General_Concept
.	O
</s>
<s>
Word	B-General_Concept
embedding	I-General_Concept
,	O
such	O
as	O
word2vec	B-Algorithm
,	O
can	O
be	O
thought	O
of	O
as	O
a	O
representational	O
layer	O
in	O
a	O
deep	B-Algorithm
learning	I-Algorithm
architecture	O
that	O
transforms	O
an	O
atomic	O
word	O
into	O
a	O
positional	O
representation	O
of	O
the	O
word	O
relative	O
to	O
other	O
words	O
in	O
the	O
dataset	O
;	O
the	O
position	O
is	O
represented	O
as	O
a	O
point	O
in	O
a	O
vector	O
space	O
.	O
</s>
<s>
Using	O
word	B-General_Concept
embedding	I-General_Concept
as	O
an	O
RNN	B-Algorithm
input	O
layer	O
allows	O
the	O
network	O
to	O
parse	O
sentences	O
and	O
phrases	O
using	O
an	O
effective	O
compositional	O
vector	O
grammar	O
.	O
</s>
<s>
A	O
compositional	O
vector	O
grammar	O
can	O
be	O
thought	O
of	O
as	O
probabilistic	B-General_Concept
context	I-General_Concept
free	I-General_Concept
grammar	I-General_Concept
(	O
PCFG	B-General_Concept
)	O
implemented	O
by	O
an	O
RNN	B-Algorithm
.	O
</s>
<s>
Recursive	O
auto-encoders	B-Algorithm
built	O
atop	O
word	B-General_Concept
embeddings	I-General_Concept
can	O
assess	O
sentence	O
similarity	O
and	O
detect	O
paraphrasing	O
.	O
</s>
<s>
Deep	O
neural	O
architectures	O
provide	O
the	O
best	O
results	O
for	O
constituency	B-General_Concept
parsing	I-General_Concept
,	O
sentiment	O
analysis	O
,	O
information	O
retrieval	O
,	O
spoken	O
language	O
understanding	O
,	O
machine	B-Application
translation	I-Application
,	O
contextual	O
entity	O
linking	O
,	O
writing	O
style	O
recognition	O
,	O
Text	O
classification	O
and	O
others	O
.	O
</s>
<s>
Recent	O
developments	O
generalize	O
word	B-General_Concept
embedding	I-General_Concept
to	O
sentence	B-General_Concept
embedding	I-General_Concept
.	O
</s>
<s>
Google	B-Application
Translate	I-Application
(	O
GT	O
)	O
uses	O
a	O
large	O
end-to-end	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
network	O
.	O
</s>
<s>
Google	B-Algorithm
Neural	I-Algorithm
Machine	I-Algorithm
Translation	I-Algorithm
(	O
GNMT	B-Algorithm
)	O
uses	O
an	O
example-based	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
method	O
in	O
which	O
the	O
system	O
"	O
learns	O
from	O
millions	O
of	O
examples.	O
"	O
</s>
<s>
Google	B-Application
Translate	I-Application
supports	O
over	O
one	O
hundred	O
languages	O
.	O
</s>
<s>
Research	O
has	O
explored	O
use	O
of	O
deep	B-Algorithm
learning	I-Algorithm
to	O
predict	O
the	O
biomolecular	O
targets	O
,	O
off-targets	O
,	O
and	O
toxic	O
effects	O
of	O
environmental	O
chemicals	O
in	O
nutrients	O
,	O
household	O
products	O
and	O
drugs	O
.	O
</s>
<s>
AtomNet	O
is	O
a	O
deep	B-Algorithm
learning	I-Algorithm
system	O
for	O
structure-based	O
rational	O
drug	O
design	O
.	O
</s>
<s>
In	O
2017	O
graph	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
were	O
used	O
for	O
the	O
first	O
time	O
to	O
predict	O
various	O
properties	O
of	O
molecules	O
in	O
a	O
large	O
toxicology	O
data	O
set	O
.	O
</s>
<s>
In	O
2019	O
,	O
generative	O
neural	B-Architecture
networks	I-Architecture
were	O
used	O
to	O
produce	O
molecules	O
that	O
were	O
validated	O
experimentally	O
all	O
the	O
way	O
into	O
mice	O
.	O
</s>
<s>
Deep	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
has	O
been	O
used	O
to	O
approximate	O
the	O
value	O
of	O
possible	O
direct	O
marketing	O
actions	O
,	O
defined	O
in	O
terms	O
of	O
RFM	O
variables	O
.	O
</s>
<s>
Recommendation	O
systems	O
have	O
used	O
deep	B-Algorithm
learning	I-Algorithm
to	O
extract	O
meaningful	O
features	O
for	O
a	O
latent	O
factor	O
model	O
for	O
content-based	O
music	O
and	O
journal	O
recommendations	O
.	O
</s>
<s>
Multi-view	O
deep	B-Algorithm
learning	I-Algorithm
has	O
been	O
applied	O
for	O
learning	O
user	O
preferences	O
from	O
multiple	O
domains	O
.	O
</s>
<s>
An	O
autoencoder	B-Algorithm
ANN	O
was	O
used	O
in	O
bioinformatics	O
,	O
to	O
predict	O
gene	O
ontology	O
annotations	O
and	O
gene-function	O
relationships	O
.	O
</s>
<s>
In	O
medical	O
informatics	O
,	O
deep	B-Algorithm
learning	I-Algorithm
was	O
used	O
to	O
predict	O
sleep	O
quality	O
based	O
on	O
data	O
from	O
wearables	O
and	O
predictions	O
of	O
health	O
complications	O
from	O
electronic	B-Application
health	I-Application
record	I-Application
data	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
has	O
been	O
shown	O
to	O
produce	O
competitive	O
results	O
in	O
medical	O
application	O
such	O
as	O
cancer	O
cell	O
classification	O
,	O
lesion	O
detection	O
,	O
organ	O
segmentation	O
and	O
image	O
enhancement	O
.	O
</s>
<s>
Modern	O
deep	B-Algorithm
learning	I-Algorithm
tools	O
demonstrate	O
the	O
high	O
accuracy	O
of	O
detecting	O
various	O
diseases	O
and	O
the	O
helpfulness	O
of	O
their	O
use	O
by	O
specialists	O
to	O
improve	O
the	O
diagnosis	O
efficiency	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
has	O
been	O
used	O
to	O
interpret	O
large	O
,	O
many-dimensioned	O
advertising	O
datasets	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
has	O
been	O
successfully	O
applied	O
to	O
inverse	O
problems	O
such	O
as	O
denoising	O
,	O
super-resolution	B-Algorithm
,	O
inpainting	B-Algorithm
,	O
and	O
film	O
colorization	O
.	O
</s>
<s>
These	O
applications	O
include	O
learning	O
methods	O
such	O
as	O
"	O
Shrinkage	O
Fields	O
for	O
Effective	O
Image	O
Restoration	O
"	O
which	O
trains	O
on	O
an	O
image	O
dataset	O
,	O
and	O
Deep	B-Algorithm
Image	I-Algorithm
Prior	I-Algorithm
,	O
which	O
trains	O
on	O
the	O
image	O
that	O
needs	O
restoration	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
being	O
successfully	O
applied	O
to	O
financial	O
fraud	O
detection	O
,	O
tax	O
evasion	O
detection	O
,	O
and	O
anti-money	O
laundering	O
.	O
</s>
<s>
The	O
United	O
States	O
Department	O
of	O
Defense	O
applied	O
deep	B-Algorithm
learning	I-Algorithm
to	O
train	O
robots	O
in	O
new	O
tasks	O
through	O
observation	O
.	O
</s>
<s>
Physics	O
informed	O
neural	B-Architecture
networks	I-Architecture
have	O
been	O
used	O
to	O
solve	O
partial	O
differential	O
equations	O
in	O
both	O
forward	O
and	O
inverse	O
problems	O
in	O
a	O
data	O
driven	O
manner	O
.	O
</s>
<s>
Using	O
physics	O
informed	O
neural	B-Architecture
networks	I-Architecture
does	O
not	O
require	O
the	O
often	O
expensive	O
mesh	O
generation	O
that	O
conventional	O
CFD	O
methods	O
relies	O
on	O
.	O
</s>
<s>
Several	O
works	O
showed	O
the	O
better	O
and	O
superior	O
performance	O
of	O
the	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
compared	O
to	O
analytical	O
methods	O
for	O
various	O
applications	O
,	O
e.g.	O
,	O
spectral	O
imaging	O
and	O
ultrasound	O
imaging	O
.	O
</s>
<s>
used	O
deep	O
neural	B-Architecture
networks	I-Architecture
to	O
train	O
an	O
epigenetic	O
aging	O
clock	O
of	O
unprecedented	O
accuracy	O
using	O
>6	O
,	O
000	O
blood	O
samples	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
is	O
closely	O
related	O
to	O
a	O
class	O
of	O
theories	O
of	O
brain	O
development	O
(	O
specifically	O
,	O
neocortical	O
development	O
)	O
proposed	O
by	O
cognitive	O
neuroscientists	O
in	O
the	O
early	O
1990s	O
.	O
</s>
<s>
These	O
developmental	O
theories	O
were	O
instantiated	O
in	O
computational	O
models	O
,	O
making	O
them	O
predecessors	O
of	O
deep	B-Algorithm
learning	I-Algorithm
systems	O
.	O
</s>
<s>
These	O
developmental	O
models	O
share	O
the	O
property	O
that	O
various	O
proposed	O
learning	O
dynamics	O
in	O
the	O
brain	O
(	O
e.g.	O
,	O
a	O
wave	O
of	O
nerve	O
growth	O
factor	O
)	O
support	O
the	O
self-organization	O
somewhat	O
analogous	O
to	O
the	O
neural	B-Architecture
networks	I-Architecture
utilized	O
in	O
deep	B-Algorithm
learning	I-Algorithm
models	O
.	O
</s>
<s>
Like	O
the	O
neocortex	O
,	O
neural	B-Architecture
networks	I-Architecture
employ	O
a	O
hierarchy	O
of	O
layered	O
filters	O
in	O
which	O
each	O
layer	O
considers	O
information	O
from	O
a	O
prior	O
layer	O
(	O
or	O
the	O
operating	O
environment	O
)	O
,	O
and	O
then	O
passes	O
its	O
output	O
(	O
and	O
possibly	O
the	O
original	O
input	O
)	O
,	O
to	O
other	O
layers	O
.	O
</s>
<s>
This	O
process	O
yields	O
a	O
self-organizing	O
stack	O
of	O
transducers	B-Algorithm
,	O
well-tuned	O
to	O
their	O
operating	O
environment	O
.	O
</s>
<s>
A	O
variety	O
of	O
approaches	O
have	O
been	O
used	O
to	O
investigate	O
the	O
plausibility	O
of	O
deep	B-Algorithm
learning	I-Algorithm
models	O
from	O
a	O
neurobiological	O
perspective	O
.	O
</s>
<s>
On	O
the	O
one	O
hand	O
,	O
several	O
variants	O
of	O
the	O
backpropagation	B-Algorithm
algorithm	O
have	O
been	O
proposed	O
in	O
order	O
to	O
increase	O
its	O
processing	O
realism	O
.	O
</s>
<s>
Other	O
researchers	O
have	O
argued	O
that	O
unsupervised	B-General_Concept
forms	O
of	O
deep	B-Algorithm
learning	I-Algorithm
,	O
such	O
as	O
those	O
based	O
on	O
hierarchical	O
generative	O
models	O
and	O
deep	B-Algorithm
belief	I-Algorithm
networks	I-Algorithm
,	O
may	O
be	O
closer	O
to	O
biological	O
reality	O
.	O
</s>
<s>
In	O
this	O
respect	O
,	O
generative	O
neural	B-Architecture
network	I-Architecture
models	I-Architecture
have	O
been	O
related	O
to	O
neurobiological	O
evidence	O
about	O
sampling-based	O
processing	O
in	O
the	O
cerebral	O
cortex	O
.	O
</s>
<s>
For	O
example	O
,	O
the	O
computations	O
performed	O
by	O
deep	B-Algorithm
learning	I-Algorithm
units	O
could	O
be	O
similar	O
to	O
those	O
of	O
actual	O
neurons	O
and	O
neural	O
populations	O
.	O
</s>
<s>
Similarly	O
,	O
the	O
representations	O
developed	O
by	O
deep	B-Algorithm
learning	I-Algorithm
models	O
are	O
similar	O
to	O
those	O
measured	O
in	O
the	O
primate	O
visual	O
system	O
both	O
at	O
the	O
single-unit	O
and	O
at	O
the	O
population	O
levels	O
.	O
</s>
<s>
Facebook	B-Application
's	O
AI	B-Application
lab	O
performs	O
tasks	O
such	O
as	O
automatically	B-Application
tagging	I-Application
uploaded	I-Application
pictures	I-Application
with	O
the	O
names	O
of	O
the	O
people	O
in	O
them	O
.	O
</s>
<s>
Google	B-Application
's	I-Application
DeepMind	B-Application
Technologies	I-Application
developed	O
a	O
system	O
capable	O
of	O
learning	O
how	O
to	O
play	O
Atari	O
video	O
games	O
using	O
only	O
pixels	O
as	O
data	O
input	O
.	O
</s>
<s>
In	O
2015	O
they	O
demonstrated	O
their	O
AlphaGo	B-Application
system	O
,	O
which	O
learned	O
the	O
game	O
of	O
Go	O
well	O
enough	O
to	O
beat	O
a	O
professional	O
Go	O
player	O
.	O
</s>
<s>
Google	B-Application
Translate	I-Application
uses	O
a	O
neural	B-Architecture
network	I-Architecture
to	O
translate	O
between	O
more	O
than	O
100	O
languages	O
.	O
</s>
<s>
In	O
2017	O
,	O
Covariant.ai	O
was	O
launched	O
,	O
which	O
focuses	O
on	O
integrating	O
deep	B-Algorithm
learning	I-Algorithm
into	O
factories	O
.	O
</s>
<s>
Deep	O
TAMER	O
used	O
deep	B-Algorithm
learning	I-Algorithm
to	O
provide	O
a	O
robot	O
the	O
ability	O
to	O
learn	O
new	O
tasks	O
through	O
observation	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
has	O
attracted	O
both	O
criticism	O
and	O
comment	O
,	O
in	O
some	O
cases	O
from	O
outside	O
the	O
field	O
of	O
computer	O
science	O
.	O
</s>
<s>
Learning	O
in	O
the	O
most	O
common	O
deep	O
architectures	O
is	O
implemented	O
using	O
well-understood	O
gradient	B-Algorithm
descent	I-Algorithm
.	O
</s>
<s>
However	O
,	O
the	O
theory	O
surrounding	O
other	O
algorithms	O
,	O
such	O
as	O
contrastive	B-Algorithm
divergence	I-Algorithm
is	O
less	O
clear	O
.	O
</s>
<s>
Deep	B-Algorithm
learning	I-Algorithm
methods	O
are	O
often	O
looked	O
at	O
as	O
a	O
black	B-Device
box	I-Device
,	O
with	O
most	O
confirmations	O
done	O
empirically	O
,	O
rather	O
than	O
theoretically	O
.	O
</s>
<s>
Others	O
point	O
out	O
that	O
deep	B-Algorithm
learning	I-Algorithm
should	O
be	O
looked	O
at	O
as	O
a	O
step	O
towards	O
realizing	O
strong	O
AI	B-Application
,	O
not	O
as	O
an	O
all-encompassing	O
solution	O
.	O
</s>
<s>
Despite	O
the	O
power	O
of	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
,	O
they	O
still	O
lack	O
much	O
of	O
the	O
functionality	O
needed	O
for	O
realizing	O
this	O
goal	O
entirely	O
.	O
</s>
<s>
Research	O
psychologist	O
Gary	O
Marcus	O
noted	O
:	O
"	O
Realistically	O
,	O
deep	B-Algorithm
learning	I-Algorithm
is	O
only	O
part	O
of	O
the	O
larger	O
challenge	O
of	O
building	O
intelligent	B-Application
machines	I-Application
.	O
</s>
<s>
Such	O
techniques	O
lack	O
ways	O
of	O
representing	O
causal	B-Application
relationships	I-Application
(	O
...	O
)	O
have	O
no	O
obvious	O
ways	O
of	O
performing	O
logical	O
inferences	O
,	O
and	O
they	O
are	O
also	O
still	O
a	O
long	O
way	O
from	O
integrating	O
abstract	O
knowledge	O
,	O
such	O
as	O
information	O
about	O
what	O
objects	O
are	O
,	O
what	O
they	O
are	O
for	O
,	O
and	O
how	O
they	O
are	O
typically	O
used	O
.	O
</s>
<s>
The	O
most	O
powerful	O
A.I.	B-Application
</s>
<s>
systems	O
,	O
like	O
Watson	B-Application
(	O
...	O
)	O
use	O
techniques	O
like	O
deep	B-Algorithm
learning	I-Algorithm
as	O
just	O
one	O
element	O
in	O
a	O
very	O
complicated	O
ensemble	O
of	O
techniques	O
,	O
ranging	O
from	O
the	O
statistical	O
technique	O
of	O
Bayesian	O
inference	O
to	O
deductive	O
reasoning.	O
"	O
</s>
<s>
In	O
further	O
reference	O
to	O
the	O
idea	O
that	O
artistic	O
sensitivity	O
might	O
be	O
inherent	O
in	O
relatively	O
low	O
levels	O
of	O
the	O
cognitive	O
hierarchy	O
,	O
a	O
published	O
series	O
of	O
graphic	O
representations	O
of	O
the	O
internal	O
states	O
of	O
deep	O
(	O
20-30	O
layers	O
)	O
neural	B-Architecture
networks	I-Architecture
attempting	O
to	O
discern	O
within	O
essentially	O
random	O
data	O
the	O
images	O
on	O
which	O
they	O
were	O
trained	O
demonstrate	O
a	O
visual	O
appeal	O
:	O
the	O
original	O
research	O
notice	O
received	O
well	O
over	O
1,000	O
comments	O
,	O
and	O
was	O
the	O
subject	O
of	O
what	O
was	O
for	O
a	O
time	O
the	O
most	O
frequently	O
accessed	O
article	O
on	O
The	O
Guardian	O
's	O
website	O
.	O
</s>
<s>
Furthermore	O
,	O
while	O
deep	B-Algorithm
learning	I-Algorithm
consists	O
of	O
dozens	O
and	O
even	O
hundreds	O
of	O
layers	O
,	O
the	O
brain	O
itself	O
consists	O
of	O
very	O
few	O
layers	O
.	O
</s>
<s>
Simulations	O
on	O
shallow	O
networks	O
,	O
which	O
are	O
closer	O
to	O
the	O
brain	O
dynamics	O
,	O
indicate	O
a	O
similar	O
performance	O
as	O
deep	B-Algorithm
learning	I-Algorithm
with	O
a	O
lower	O
complexity	O
.	O
</s>
<s>
Some	O
deep	B-Algorithm
learning	I-Algorithm
architectures	O
display	O
problematic	O
behaviors	O
,	O
such	O
as	O
confidently	O
classifying	O
unrecognizable	O
images	O
as	O
belonging	O
to	O
a	O
familiar	O
category	O
of	O
ordinary	O
images	O
(	O
2014	O
)	O
and	O
misclassifying	O
minuscule	O
perturbations	O
of	O
correctly	O
classified	O
images	O
(	O
2013	O
)	O
.	O
</s>
<s>
These	O
issues	O
may	O
possibly	O
be	O
addressed	O
by	O
deep	B-Algorithm
learning	I-Algorithm
architectures	O
that	O
internally	O
form	O
states	O
homologous	O
to	O
image-grammar	O
decompositions	O
of	O
observed	O
entities	O
and	O
events	O
.	O
</s>
<s>
Learning	B-Algorithm
a	I-Algorithm
grammar	I-Algorithm
(	O
visual	O
or	O
linguistic	O
)	O
from	O
training	O
data	O
would	O
be	O
equivalent	O
to	O
restricting	O
the	O
system	O
to	O
commonsense	O
reasoning	O
that	O
operates	O
on	O
concepts	O
in	O
terms	O
of	O
grammatical	O
production	O
rules	O
and	O
is	O
a	O
basic	O
goal	O
of	O
both	O
human	O
language	O
acquisition	O
and	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
.	O
</s>
<s>
As	O
deep	B-Algorithm
learning	I-Algorithm
moves	O
from	O
the	O
lab	O
into	O
the	O
world	O
,	O
research	O
and	O
experience	O
show	O
that	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
are	O
vulnerable	O
to	O
hacks	O
and	O
deception	B-Application
.	O
</s>
<s>
One	O
defense	O
is	O
reverse	O
image	O
search	O
,	O
in	O
which	O
a	O
possible	O
fake	O
image	O
is	O
submitted	O
to	O
a	O
site	O
such	O
as	O
TinEye	B-General_Concept
that	O
can	O
then	O
find	O
other	O
instances	O
of	O
it	O
.	O
</s>
<s>
ANNs	O
can	O
however	O
be	O
further	O
trained	O
to	O
detect	O
attempts	O
at	O
deception	B-Application
,	O
potentially	O
leading	O
attackers	O
and	O
defenders	O
into	O
an	O
arms	O
race	O
similar	O
to	O
the	O
kind	O
that	O
already	O
defines	O
the	O
malware	O
defense	O
industry	O
.	O
</s>
<s>
ANNs	O
have	O
been	O
trained	O
to	O
defeat	O
ANN-based	O
anti-malware	O
software	O
by	O
repeatedly	O
attacking	O
a	O
defense	O
with	O
malware	O
that	O
was	O
continually	O
altered	O
by	O
a	O
genetic	B-Algorithm
algorithm	I-Algorithm
until	O
it	O
tricked	O
the	O
anti-malware	O
while	O
retaining	O
its	O
ability	O
to	O
damage	O
the	O
target	O
.	O
</s>
<s>
In	O
2016	O
,	O
another	O
group	O
demonstrated	O
that	O
certain	O
sounds	O
could	O
make	O
the	B-Application
Google	I-Application
Now	O
voice	B-Application
command	I-Application
system	O
open	O
a	O
particular	O
web	O
address	O
,	O
and	O
hypothesized	O
that	O
this	O
could	O
"	O
serve	O
as	O
a	O
stepping	O
stone	O
for	O
further	O
attacks	O
(	O
e.g.	O
,	O
opening	O
a	O
web	O
page	O
hosting	O
drive-by	O
malware	O
)	O
.	O
"	O
</s>
<s>
Most	O
Deep	B-Algorithm
Learning	I-Algorithm
systems	O
rely	O
on	O
training	O
and	O
verification	O
data	O
that	O
is	O
generated	O
and/or	O
annotated	O
by	O
humans	O
.	O
</s>
<s>
It	O
has	O
been	O
argued	O
in	O
media	O
philosophy	O
that	O
not	O
only	O
low-paid	O
clickwork	B-Application
(	O
e.g.	O
</s>
<s>
CAPTCHAs	O
for	O
image	O
recognition	O
or	O
click-tracking	O
on	O
Google	B-Application
search	O
results	O
pages	O
)	O
,	O
(	O
3	O
)	O
exploitation	O
of	O
social	O
motivations	O
(	O
e.g.	O
</s>
<s>
tagging	O
faces	O
on	O
Facebook	B-Application
to	O
obtain	O
labeled	B-General_Concept
facial	O
images	O
)	O
,	O
(	O
4	O
)	O
information	B-Application
mining	I-Application
(	O
e.g.	O
</s>
<s>
by	O
leveraging	O
quantified-self	O
devices	O
such	O
as	O
activity	B-Device
trackers	I-Device
)	O
and	O
(	O
5	O
)	O
clickwork	B-Application
.	O
</s>
<s>
Mühlhoff	O
argues	O
that	O
in	O
most	O
commercial	O
end-user	O
applications	O
of	O
Deep	B-Algorithm
Learning	I-Algorithm
such	O
as	O
Facebook	B-Application
's	O
face	O
recognition	O
system	O
,	O
the	O
need	O
for	O
training	O
data	O
does	O
not	O
stop	O
once	O
an	O
ANN	O
is	O
trained	O
.	O
</s>
<s>
For	O
this	O
purpose	O
Facebook	B-Application
introduced	O
the	O
feature	O
that	O
once	O
a	O
user	O
is	O
automatically	O
recognized	O
in	O
an	O
image	O
,	O
they	O
receive	O
a	O
notification	O
.	O
</s>
<s>
They	O
can	O
choose	O
whether	O
of	O
not	O
they	O
like	O
to	O
be	O
publicly	O
labeled	B-General_Concept
on	O
the	O
image	O
,	O
or	O
tell	O
Facebook	B-Application
that	O
it	O
is	O
not	O
them	O
in	O
the	O
picture	O
.	O
</s>
<s>
As	O
Mühlhoff	O
argues	O
,	O
involvement	O
of	O
human	O
users	O
to	O
generate	O
training	O
and	O
verification	O
data	O
is	O
so	O
typical	O
for	O
most	O
commercial	O
end-user	O
applications	O
of	O
Deep	B-Algorithm
Learning	I-Algorithm
that	O
such	O
systems	O
may	O
be	O
referred	O
to	O
as	O
"	O
human-aided	O
artificial	B-Application
intelligence	I-Application
"	O
.	O
</s>
