<s>
LeNet	B-Algorithm
is	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
structure	O
proposed	O
by	O
LeCun	O
et	O
al	O
.	O
</s>
<s>
In	O
general	O
,	O
LeNet	B-Algorithm
refers	O
to	O
LeNet-5	B-Algorithm
and	O
is	O
a	O
simple	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
.	O
</s>
<s>
Convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
are	O
a	O
kind	O
of	O
feed-forward	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
whose	O
artificial	O
neurons	O
can	O
respond	O
to	O
a	O
part	O
of	O
the	O
surrounding	O
cells	O
in	O
the	O
coverage	O
range	O
and	O
perform	O
well	O
in	O
large-scale	O
image	O
processing	O
.	O
</s>
<s>
LeNet-5	B-Algorithm
was	O
one	O
of	O
the	O
earliest	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
and	O
promoted	O
the	O
development	O
of	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Since	O
1988	O
,	O
after	O
years	O
of	O
research	O
and	O
many	O
successful	O
iterations	O
,	O
the	O
pioneering	O
work	O
has	O
been	O
named	O
LeNet-5	B-Algorithm
.	O
</s>
<s>
at	O
Bell	O
Labs	O
first	O
applied	O
the	O
backpropagation	B-Algorithm
algorithm	I-Algorithm
to	O
practical	O
applications	O
,	O
and	O
believed	O
that	O
the	O
ability	O
to	O
learn	O
network	O
generalization	O
could	O
be	O
greatly	O
enhanced	O
by	O
providing	O
constraints	O
from	O
the	O
task	O
's	O
domain	O
.	O
</s>
<s>
He	O
combined	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
trained	O
by	O
backpropagation	B-Algorithm
algorithms	O
to	O
read	O
handwritten	O
numbers	O
and	O
successfully	O
applied	O
it	O
in	O
identifying	O
handwritten	O
zip	B-Language
code	O
numbers	O
provided	O
by	O
the	O
US	O
Postal	O
Service	O
.	O
</s>
<s>
This	O
was	O
the	O
prototype	O
of	O
what	O
later	O
came	O
to	O
be	O
called	O
LeNet	B-Algorithm
.	O
</s>
<s>
In	O
1990	O
,	O
their	O
paper	O
described	O
the	O
application	O
of	O
backpropagation	B-Algorithm
networks	O
in	O
handwritten	O
digit	O
recognition	O
again	O
.	O
</s>
<s>
While	O
the	O
architecture	O
of	O
the	O
best	O
performing	O
neural	O
networks	O
today	O
are	O
not	O
the	O
same	O
as	O
that	O
of	O
LeNet	B-Algorithm
,	O
the	O
network	O
was	O
the	O
starting	O
point	O
for	O
a	O
large	O
number	O
of	O
neural	O
network	O
architectures	O
,	O
and	O
also	O
brought	O
inspiration	O
to	O
the	O
field	O
.	O
</s>
<s>
Backpropagation	B-Algorithm
applied	O
to	O
handwritten	O
zip	B-Language
code	O
recognition	O
.	O
</s>
<s>
Technical	O
Report	O
CRG-TR-89-4	O
,	O
Department	O
of	O
Computer	O
Science	O
,	O
University	O
of	O
Toronto.1990Their	O
paper	O
describes	O
the	O
application	O
of	O
backpropagation	B-Algorithm
networks	O
in	O
handwritten	O
digit	O
recognition	O
once	O
againLeCun	O
,	O
Y.	O
;	O
Boser	O
,	O
B.	O
;	O
Denker	O
,	O
J	O
.	O
S.	O
;	O
Henderson	O
,	O
D.	O
;	O
Howard	O
,	O
R	O
.	O
E.	O
;	O
Hubbard	O
,	O
W	O
.	O
&	O
Jackel	O
,	O
L	O
.	O
D	O
.	O
(	O
1990	O
)	O
.	O
</s>
<s>
Handwritten	O
digit	O
recognition	O
with	O
a	O
back-propagation	B-Algorithm
network	O
.	O
</s>
<s>
The	O
results	O
show	O
that	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
outperform	O
all	O
other	O
models.LeCun	O
,	O
Y.	O
;	O
Bottou	O
,	O
L.	O
;	O
Bengio	O
,	O
Y	O
.	O
</s>
<s>
As	O
a	O
representative	O
of	O
the	O
early	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
,	O
LeNet	B-Algorithm
possesses	O
the	O
basic	O
units	O
of	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
,	O
such	O
as	O
convolutional	B-Architecture
layer	I-Architecture
,	O
pooling	O
layer	O
and	O
full	O
connection	O
layer	O
,	O
laying	O
a	O
foundation	O
for	O
the	O
future	O
development	O
of	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
.	O
</s>
<s>
As	O
shown	O
in	O
the	O
figure	O
(	O
input	O
image	O
data	O
with	O
32*32	O
pixels	O
)	O
:	O
LeNet-5	B-Algorithm
consists	O
of	O
seven	O
layers	O
.	O
</s>
<s>
In	O
the	O
figure	O
,	O
Cx	O
represents	O
convolution	B-Language
layer	I-Language
,	O
Sx	O
represents	O
sub-sampling	O
layer	O
,	O
Fx	O
represents	O
complete	O
connection	O
layer	O
,	O
and	O
x	O
represents	O
layer	O
index	O
.	O
</s>
<s>
Layer	O
C1	O
is	O
a	O
convolution	B-Language
layer	I-Language
with	O
six	O
convolution	B-Language
kernels	O
of	O
5x5	O
and	O
the	O
size	O
of	O
feature	O
mapping	O
is	O
28x28	O
,	O
which	O
can	O
prevent	O
the	O
information	O
of	O
the	O
input	O
image	O
from	O
falling	O
out	O
of	O
the	O
boundary	O
of	O
convolution	B-Language
kernel	O
.	O
</s>
<s>
Layer	O
C3	O
is	O
a	O
convolution	B-Language
layer	I-Language
with	O
16	O
5-5	O
convolution	B-Language
kernels	O
.	O
</s>
<s>
Layer	O
C5	O
is	O
a	O
convolution	B-Language
layer	I-Language
with	O
120	O
convolution	B-Language
kernels	O
of	O
size	O
5x5	O
.	O
</s>
<s>
C5	O
is	O
labeled	O
as	O
a	O
convolutional	B-Architecture
layer	I-Architecture
instead	O
of	O
a	O
fully	O
connected	O
layer	O
,	O
because	O
if	O
LeNet-5	B-Algorithm
input	O
becomes	O
larger	O
and	O
its	O
structure	O
remains	O
unchanged	O
,	O
its	O
output	O
size	O
will	O
be	O
greater	O
than	O
1x1	O
,	O
i.e.	O
</s>
<s>
Recognizing	O
simple	O
digit	O
images	O
is	O
the	O
most	O
classic	O
application	O
of	O
LeNet	B-Algorithm
as	O
it	O
was	O
created	O
because	O
of	O
that	O
.	O
</s>
<s>
created	O
the	O
initial	O
form	O
of	O
LeNet	B-Algorithm
in	O
1989	O
.	O
</s>
<s>
The	O
paper	O
Backpropagation	B-Algorithm
Applied	O
to	O
Handwritten	O
Zip	B-Language
Code	O
Recognition	O
demonstrates	O
how	O
such	O
constraints	O
can	O
be	O
integrated	O
into	O
a	O
backpropagation	B-Algorithm
network	O
through	O
the	O
architecture	O
of	O
the	O
network	O
.	O
</s>
<s>
And	O
it	O
had	O
been	O
successfully	O
applied	O
to	O
the	O
recognition	O
of	O
handwritten	O
zip	B-Language
code	O
digits	O
provided	O
by	O
the	O
U.S.	O
</s>
<s>
The	O
LeNet-5	B-Algorithm
means	O
the	O
emergence	O
of	O
CNN	B-Architecture
and	O
defines	O
the	O
basic	O
components	O
of	O
CNN	B-Architecture
.	O
</s>
<s>
But	O
it	O
was	O
not	O
popular	O
at	O
that	O
time	O
because	O
of	O
the	O
lack	O
of	O
hardware	O
equipment	O
,	O
especially	O
GPU	B-Architecture
(	O
Graphics	B-Architecture
Processing	I-Architecture
Unit	I-Architecture
,	O
a	O
specialized	O
electronic	O
circuit	O
designed	O
to	O
rapidly	O
manipulate	O
and	O
alter	O
memory	B-General_Concept
to	O
accelerate	O
the	O
creation	O
of	O
images	O
in	O
a	O
frame	B-Algorithm
buffer	I-Algorithm
intended	O
for	O
output	O
to	O
a	O
display	B-Device
device	I-Device
)	O
and	O
other	O
algorithm	O
,	O
such	O
as	O
SVM	B-Algorithm
can	O
achieve	O
similar	O
effects	O
or	O
even	O
exceed	O
the	O
LeNet	B-Algorithm
.	O
</s>
<s>
Since	O
the	O
success	O
of	O
AlexNet	O
in	O
2012	O
,	O
CNN	B-Architecture
has	O
become	O
the	O
best	O
choice	O
for	O
computer	O
vision	O
applications	O
and	O
many	O
different	O
types	O
of	O
CNN	B-Architecture
has	O
been	O
created	O
,	O
such	O
as	O
the	O
R-CNN	O
series	O
.	O
</s>
<s>
Nowadays	O
,	O
CNN	B-Architecture
models	O
are	O
quite	O
different	O
from	O
LeNet	B-Algorithm
,	O
but	O
they	O
are	O
all	O
developed	O
on	O
the	O
basis	O
of	O
LeNet	B-Algorithm
.	O
</s>
<s>
Recently	O
a	O
three	O
layer	O
tree	O
architecture	O
imitating	O
LeNet-5	B-Algorithm
and	O
consisting	O
of	O
only	O
one	O
convolutional	B-Architecture
layer	I-Architecture
,	O
has	O
achieved	O
a	O
similar	O
success	O
rate	O
on	O
the	O
CIFAR-10	O
dataset	O
.	O
</s>
<s>
Increasing	O
the	O
number	O
of	O
filters	O
for	O
the	O
LeNet	B-Algorithm
architecture	O
results	O
in	O
a	O
power	O
law	O
decay	O
of	O
the	O
error	O
rate	O
.	O
</s>
<s>
These	O
results	O
indicate	O
that	O
a	O
shallow	O
network	O
can	O
achieve	O
the	O
same	O
performance	O
as	O
deep	B-Algorithm
learning	I-Algorithm
architectures	O
.	O
</s>
