<s>
The	O
MNIST	B-General_Concept
database	I-General_Concept
(	O
Modified	O
National	O
Institute	O
of	O
Standards	O
and	O
Technology	O
database	O
)	O
is	O
a	O
large	O
database	O
of	O
handwritten	O
digits	O
that	O
is	O
commonly	O
used	O
for	O
training	O
various	O
image	B-Algorithm
processing	I-Algorithm
systems	O
.	O
</s>
<s>
Furthermore	O
,	O
the	O
black	O
and	O
white	O
images	O
from	O
NIST	O
were	O
normalized	B-Algorithm
to	O
fit	O
into	O
a	O
28x28	O
pixel	O
bounding	O
box	O
and	O
anti-aliased	B-Algorithm
,	O
which	O
introduced	O
grayscale	O
levels	O
.	O
</s>
<s>
The	O
MNIST	B-General_Concept
database	I-General_Concept
contains	O
60,000	O
training	O
images	O
and	O
10,000	O
testing	O
images	O
.	O
</s>
<s>
In	O
their	O
original	O
paper	O
,	O
they	O
use	O
a	O
support-vector	B-Algorithm
machine	I-Algorithm
to	O
get	O
an	O
error	O
rate	O
of	O
0.8	O
%	O
.	O
</s>
<s>
Extended	O
MNIST	B-General_Concept
(	O
EMNIST	O
)	O
is	O
a	O
newer	O
dataset	O
developed	O
and	O
released	O
by	O
NIST	O
to	O
be	O
the	O
(	O
final	O
)	O
successor	O
to	O
MNIST	B-General_Concept
.	O
</s>
<s>
MNIST	B-General_Concept
included	O
images	O
only	O
of	O
handwritten	O
digits	O
.	O
</s>
<s>
The	O
images	O
in	O
EMNIST	O
were	O
converted	O
into	O
the	O
same	O
28x28	O
pixel	O
format	O
,	O
by	O
the	O
same	O
process	O
,	O
as	O
were	O
the	O
MNIST	B-General_Concept
images	O
.	O
</s>
<s>
Accordingly	O
,	O
tools	O
which	O
work	O
with	O
the	O
older	O
,	O
smaller	O
,	O
MNIST	B-General_Concept
dataset	I-General_Concept
will	O
likely	O
work	O
unmodified	O
with	O
EMNIST	O
.	O
</s>
<s>
The	O
set	O
of	O
images	O
in	O
the	O
MNIST	B-General_Concept
database	I-General_Concept
was	O
created	O
in	O
1998	O
as	O
a	O
combination	O
of	O
two	O
of	O
NIST	O
's	O
databases	O
:	O
Special	O
Database	O
1	O
and	O
Special	O
Database	O
3	O
.	O
</s>
<s>
Some	O
researchers	O
have	O
achieved	O
"	O
near-human	O
performance	O
"	O
on	O
the	O
MNIST	B-General_Concept
database	I-General_Concept
,	O
using	O
a	O
committee	O
of	O
neural	B-Architecture
networks	I-Architecture
;	O
in	O
the	O
same	O
paper	O
,	O
the	O
authors	O
achieve	O
performance	O
double	O
that	O
of	O
humans	O
on	O
other	O
recognition	O
tasks	O
.	O
</s>
<s>
The	O
highest	O
error	O
rate	O
listed	O
on	O
the	O
original	O
website	O
of	O
the	O
database	O
is	O
12	O
percent	O
,	O
which	O
is	O
achieved	O
using	O
a	O
simple	O
linear	B-General_Concept
classifier	I-General_Concept
with	O
no	O
preprocessing	B-General_Concept
.	O
</s>
<s>
The	O
systems	O
in	O
these	O
cases	O
are	O
usually	O
neural	B-Architecture
networks	I-Architecture
and	O
the	O
distortions	O
used	O
tend	O
to	O
be	O
either	O
affine	B-Algorithm
distortions	I-Algorithm
or	O
elastic	O
distortions	O
.	O
</s>
<s>
In	O
2011	O
,	O
an	O
error	O
rate	O
of	O
0.27	O
percent	O
,	O
improving	O
on	O
the	O
previous	O
best	O
result	O
,	O
was	O
reported	O
by	O
researchers	O
using	O
a	O
similar	O
system	O
of	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
In	O
2013	O
,	O
an	O
approach	O
based	O
on	O
regularization	O
of	O
neural	B-Architecture
networks	I-Architecture
using	O
DropConnect	O
has	O
been	O
claimed	O
to	O
achieve	O
a	O
0.21	O
percent	O
error	O
rate	O
.	O
</s>
<s>
In	O
2016	O
,	O
the	O
single	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
best	O
performance	O
was	O
0.25	O
percent	O
error	O
rate	O
.	O
</s>
<s>
As	O
of	O
August	O
2018	O
,	O
the	O
best	O
performance	O
of	O
a	O
single	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
trained	O
on	O
MNIST	B-General_Concept
training	O
data	O
using	O
no	O
data	B-General_Concept
augmentation	I-General_Concept
is	O
0.25	O
percent	O
error	O
rate	O
.	O
</s>
<s>
Also	O
,	O
the	O
Parallel	O
Computing	O
Center	O
(	O
Khmelnytskyi	O
,	O
Ukraine	O
)	O
obtained	O
an	O
ensemble	O
of	O
only	O
5	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
which	O
performs	O
on	O
MNIST	B-General_Concept
at	O
0.21	O
percent	O
error	O
rate	O
.	O
</s>
<s>
In	O
2018	O
,	O
researchers	O
from	O
Department	O
of	O
System	O
and	O
Information	O
Engineering	O
,	O
University	O
of	O
Virginia	O
announced	O
0.18	O
%	O
error	O
with	O
simultaneous	O
stacked	O
three	O
kind	O
of	O
neural	B-Architecture
networks	I-Architecture
(	O
fully	O
connected	O
,	O
recurrent	O
and	O
convolution	O
neural	B-Architecture
networks	I-Architecture
)	O
.	O
</s>
