<s>
Long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
used	O
in	O
the	O
fields	O
of	O
artificial	B-Application
intelligence	I-Application
and	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Unlike	O
standard	O
feedforward	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
LSTM	B-Algorithm
has	O
feedback	O
connections	O
.	O
</s>
<s>
Such	O
a	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
(	O
RNN	O
)	O
can	O
process	O
not	O
only	O
single	O
data	O
points	O
(	O
such	O
as	O
images	O
)	O
,	O
but	O
also	O
entire	O
sequences	O
of	O
data	O
(	O
such	O
as	O
speech	O
or	O
video	O
)	O
.	O
</s>
<s>
This	O
characteristic	O
makes	O
LSTM	B-Algorithm
networks	O
ideal	O
for	O
processing	B-General_Concept
and	O
predicting	O
data	O
.	O
</s>
<s>
For	O
example	O
,	O
LSTM	B-Algorithm
is	O
applicable	O
to	O
tasks	O
such	O
as	O
unsegmented	O
,	O
connected	O
handwriting	B-Application
recognition	I-Application
,	O
speech	B-Application
recognition	I-Application
,	O
machine	B-Application
translation	I-Application
,	O
speech	O
activity	O
detection	O
,	O
robot	B-Algorithm
control	O
,	O
video	O
games	O
,	O
and	O
healthcare	O
.	O
</s>
<s>
The	O
name	O
of	O
LSTM	B-Algorithm
refers	O
to	O
the	O
analogy	O
that	O
a	O
standard	O
RNN	O
has	O
both	O
"	O
long-term	O
memory	O
"	O
and	O
"	O
short-term	O
memory	O
"	O
.	O
</s>
<s>
The	O
LSTM	B-Algorithm
architecture	O
aims	O
to	O
provide	O
a	O
short-term	O
memory	O
for	O
RNN	O
that	O
can	O
last	O
thousands	O
of	O
timesteps	O
,	O
thus	O
"	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
"	O
.	O
</s>
<s>
A	O
common	O
LSTM	B-Algorithm
unit	O
is	O
composed	O
of	O
a	O
cell	O
,	O
an	O
input	O
gate	O
,	O
an	O
output	O
gate	O
and	O
a	O
forget	O
gate	O
.	O
</s>
<s>
Selectively	O
outputting	O
relevant	O
information	O
from	O
the	O
current	O
state	O
allows	O
the	O
LSTM	B-Algorithm
network	O
to	O
maintain	O
useful	O
,	O
long-term	O
dependencies	O
to	O
make	O
predictions	O
,	O
both	O
in	O
current	O
and	O
future	O
time-steps	O
.	O
</s>
<s>
LSTM	B-Algorithm
networks	O
are	O
well-suited	O
to	O
classifying	B-General_Concept
,	O
processing	B-General_Concept
and	O
making	O
predictions	O
based	O
on	O
time	O
series	O
data	O
,	O
since	O
there	O
can	O
be	O
lags	O
of	O
unknown	O
duration	O
between	O
important	O
events	O
in	O
a	O
time	O
series	O
.	O
</s>
<s>
LSTMs	B-Algorithm
were	O
developed	O
to	O
deal	O
with	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
that	O
can	O
be	O
encountered	O
when	O
training	O
traditional	O
RNNs	B-Algorithm
.	O
</s>
<s>
Relative	O
insensitivity	O
to	O
gap	O
length	O
is	O
an	O
advantage	O
of	O
LSTM	B-Algorithm
over	O
RNNs	B-Algorithm
,	O
hidden	O
Markov	O
models	O
and	O
other	O
sequence	O
learning	O
methods	O
in	O
numerous	O
applications	O
.	O
</s>
<s>
In	O
theory	O
,	O
classic	O
(	O
or	O
"	O
vanilla	O
"	O
)	O
RNNs	B-Algorithm
can	O
keep	O
track	O
of	O
arbitrary	O
long-term	O
dependencies	O
in	O
the	O
input	O
sequences	O
.	O
</s>
<s>
The	O
problem	O
with	O
vanilla	O
RNNs	B-Algorithm
is	O
computational	O
(	O
or	O
practical	O
)	O
in	O
nature	O
:	O
when	O
training	O
a	O
vanilla	O
RNN	O
using	O
back-propagation	B-Algorithm
,	O
the	O
long-term	O
gradients	O
which	O
are	O
back-propagated	O
can	O
"	B-Algorithm
vanish	I-Algorithm
"	I-Algorithm
(	O
that	O
is	O
,	O
they	O
can	O
tend	O
to	O
zero	O
)	O
or	O
"	O
explode	O
"	O
(	O
that	O
is	O
,	O
they	O
can	O
tend	O
to	O
infinity	O
)	O
,	O
because	O
of	O
the	O
computations	O
involved	O
in	O
the	O
process	O
,	O
which	O
use	O
finite-precision	B-Algorithm
numbers	I-Algorithm
.	O
</s>
<s>
RNNs	B-Algorithm
using	O
LSTM	B-Algorithm
units	O
partially	O
solve	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
,	O
because	O
LSTM	B-Algorithm
units	O
allow	O
gradients	O
to	O
also	O
flow	O
unchanged	O
.	O
</s>
<s>
However	O
,	O
LSTM	B-Algorithm
networks	O
can	O
still	O
suffer	O
from	O
the	O
exploding	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
The	O
intuition	O
behind	O
the	O
LSTM	B-Algorithm
architecture	O
is	O
to	O
create	O
an	O
additional	O
module	O
in	O
a	O
neural	B-Architecture
network	I-Architecture
that	O
learns	O
when	O
to	O
remember	O
and	O
when	O
to	O
forget	O
pertinent	O
information	O
.	O
</s>
<s>
For	O
instance	O
,	O
in	O
the	O
context	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
the	O
network	O
can	O
learn	O
grammatical	O
dependencies	O
.	O
</s>
<s>
An	O
LSTM	B-Algorithm
might	O
process	O
the	O
sentence	O
"	O
Dave	O
,	O
as	O
a	O
result	O
of	O
his	O
controversial	O
claims	O
,	O
is	O
now	O
a	O
pariah	O
"	O
by	O
remembering	O
the	O
(	O
statistically	O
likely	O
)	O
grammatical	O
gender	O
and	O
number	O
of	O
the	O
subject	O
Dave	O
,	O
note	O
that	O
this	O
information	O
is	O
pertinent	O
for	O
the	O
pronoun	O
his	O
and	O
note	O
that	O
this	O
information	O
is	O
no	O
longer	O
important	O
after	O
the	O
verb	O
is	O
.	O
</s>
<s>
So	O
,	O
for	O
example	O
,	O
is	O
not	O
just	O
one	O
unit	O
of	O
one	O
LSTM	B-Algorithm
cell	O
,	O
but	O
contains	O
LSTM	B-Algorithm
cell	O
's	O
units	O
.	O
</s>
<s>
The	O
compact	O
forms	O
of	O
the	O
equations	O
for	O
the	O
forward	O
pass	O
of	O
an	O
LSTM	B-Algorithm
cell	O
with	O
a	O
forget	O
gate	O
are	O
:	O
</s>
<s>
:	O
sigmoid	B-Algorithm
function	I-Algorithm
.	O
</s>
<s>
:	O
hyperbolic	O
tangent	O
function	O
or	O
,	O
as	O
the	O
peephole	O
LSTM	B-Algorithm
paper	O
suggests	O
,	O
.	O
</s>
<s>
The	O
figure	O
on	O
the	O
right	O
is	O
a	O
graphical	O
representation	O
of	O
an	O
LSTM	B-Algorithm
unit	O
with	O
peephole	O
connections	O
(	O
i.e.	O
</s>
<s>
a	O
peephole	O
LSTM	B-Algorithm
)	O
.	O
</s>
<s>
Each	O
of	O
the	O
gates	O
can	O
be	O
thought	O
as	O
a	O
"	O
standard	O
"	O
neuron	O
in	O
a	O
feed-forward	O
(	O
or	O
multi-layer	O
)	O
neural	B-Architecture
network	I-Architecture
:	O
that	O
is	O
,	O
they	O
compute	O
an	O
activation	O
(	O
using	O
an	O
activation	B-Algorithm
function	I-Algorithm
)	O
of	O
a	O
weighted	O
sum	O
.	O
</s>
<s>
The	O
big	O
circles	O
containing	O
an	O
S-like	O
curve	O
represent	O
the	O
application	O
of	O
a	O
differentiable	O
function	O
(	O
like	O
the	O
sigmoid	B-Algorithm
function	I-Algorithm
)	O
to	O
a	O
weighted	O
sum	O
.	O
</s>
<s>
Peephole	O
convolutional	B-Architecture
LSTM	B-Algorithm
.	O
</s>
<s>
The	O
denotes	O
the	O
convolution	B-Language
operator	O
.	O
</s>
<s>
An	O
RNN	O
using	O
LSTM	B-Algorithm
units	O
can	O
be	O
trained	O
in	O
a	O
supervised	O
fashion	O
on	O
a	O
set	O
of	O
training	O
sequences	O
,	O
using	O
an	O
optimization	O
algorithm	O
like	O
gradient	B-Algorithm
descent	I-Algorithm
combined	O
with	O
backpropagation	B-Algorithm
through	I-Algorithm
time	I-Algorithm
to	O
compute	O
the	O
gradients	O
needed	O
during	O
the	O
optimization	O
process	O
,	O
in	O
order	O
to	O
change	O
each	O
weight	O
of	O
the	O
LSTM	B-Algorithm
network	O
in	O
proportion	O
to	O
the	O
derivative	O
of	O
the	O
error	O
(	O
at	O
the	O
output	O
layer	O
of	O
the	O
LSTM	B-Algorithm
network	O
)	O
with	O
respect	O
to	O
corresponding	O
weight	O
.	O
</s>
<s>
A	O
problem	O
with	O
using	O
gradient	B-Algorithm
descent	I-Algorithm
for	O
standard	O
RNNs	B-Algorithm
is	O
that	O
error	O
gradients	O
vanish	B-Algorithm
exponentially	O
quickly	O
with	O
the	O
size	O
of	O
the	O
time	O
lag	O
between	O
important	O
events	O
.	O
</s>
<s>
However	O
,	O
with	O
LSTM	B-Algorithm
units	O
,	O
when	O
error	O
values	O
are	O
back-propagated	O
from	O
the	O
output	O
layer	O
,	O
the	O
error	O
remains	O
in	O
the	O
LSTM	B-Algorithm
unit	O
's	O
cell	O
.	O
</s>
<s>
This	O
"	O
error	O
carousel	O
"	O
continuously	O
feeds	O
error	O
back	O
to	O
each	O
of	O
the	O
LSTM	B-Algorithm
unit	O
's	O
gates	O
,	O
until	O
they	O
learn	O
to	O
cut	O
off	O
the	O
value	O
.	O
</s>
<s>
Many	O
applications	O
use	O
stacks	O
of	O
LSTM	B-Algorithm
RNNs	B-Algorithm
and	O
train	O
them	O
by	O
connectionist	B-Algorithm
temporal	I-Algorithm
classification	I-Algorithm
(	O
CTC	O
)	O
to	O
find	O
an	O
RNN	O
weight	O
matrix	O
that	O
maximizes	O
the	O
probability	O
of	O
the	O
label	O
sequences	O
in	O
a	O
training	O
set	O
,	O
given	O
the	O
corresponding	O
input	O
sequences	O
.	O
</s>
<s>
Sometimes	O
,	O
it	O
can	O
be	O
advantageous	O
to	O
train	O
(	O
parts	O
of	O
)	O
an	O
LSTM	B-Algorithm
by	O
neuroevolution	B-Algorithm
or	O
by	O
policy	O
gradient	O
methods	O
,	O
especially	O
when	O
there	O
is	O
no	O
"	O
teacher	O
"	O
(	O
that	O
is	O
,	O
training	O
labels	O
)	O
.	O
</s>
<s>
There	O
have	O
been	O
several	O
successful	O
stories	O
of	O
training	O
,	O
in	O
a	O
non-supervised	O
fashion	O
,	O
RNNs	B-Algorithm
with	O
LSTM	B-Algorithm
units	O
.	O
</s>
<s>
In	O
2018	O
,	O
Bill	O
Gates	O
called	O
it	O
a	O
"	O
huge	O
milestone	O
in	O
advancing	O
artificial	B-Application
intelligence	I-Application
"	O
when	O
bots	O
developed	O
by	O
OpenAI	O
were	O
able	O
to	O
beat	O
humans	O
in	O
the	O
game	O
of	O
Dota	O
2	O
.	O
</s>
<s>
OpenAI	O
Five	O
consists	O
of	O
five	O
independent	O
but	O
coordinated	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
In	O
2018	O
,	O
OpenAI	O
also	O
trained	O
a	O
similar	O
LSTM	B-Algorithm
by	O
policy	O
gradients	O
to	O
control	O
a	O
human-like	O
robot	B-Algorithm
hand	O
that	O
manipulates	O
physical	O
objects	O
with	O
unprecedented	O
dexterity	O
.	O
</s>
<s>
In	O
2019	O
,	O
DeepMind	B-Application
's	O
program	O
AlphaStar	B-Application
used	O
a	O
deep	O
LSTM	B-Algorithm
core	O
to	O
excel	O
at	O
the	O
complex	O
video	O
game	O
Starcraft	B-Application
II	I-Application
.	O
</s>
<s>
Applications	O
of	O
LSTM	B-Algorithm
include	O
:	O
</s>
<s>
1995	O
:	O
"	O
Long	B-Algorithm
Short-Term	I-Algorithm
Memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
"	O
is	O
published	O
in	O
a	O
technical	O
report	O
by	O
Sepp	O
Hochreiter	O
and	O
Jürgen	O
Schmidhuber	O
.	O
</s>
<s>
1996	O
:	O
LSTM	B-Algorithm
is	O
published	O
at	O
NIPS'1996	O
,	O
a	O
peer-reviewed	O
conference	O
.	O
</s>
<s>
1997	O
:	O
The	O
main	O
LSTM	B-Algorithm
paper	O
is	O
published	O
in	O
the	O
journal	O
Neural	O
Computation	O
.	O
</s>
<s>
By	O
introducing	O
Constant	O
Error	O
Carousel	O
(	O
CEC	O
)	O
units	O
,	O
LSTM	B-Algorithm
deals	O
with	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
.	O
</s>
<s>
The	O
initial	O
version	O
of	O
LSTM	B-Algorithm
block	O
included	O
cells	O
,	O
input	O
and	O
output	O
gates	O
.	O
</s>
<s>
1999	O
:	O
Felix	O
Gers	O
and	O
his	O
advisor	O
Jürgen	O
Schmidhuber	O
and	O
Fred	O
Cummins	O
introduced	O
the	O
forget	O
gate	O
(	O
also	O
called	O
"	O
keep	O
gate	O
"	O
)	O
into	O
the	O
LSTM	B-Algorithm
architecture	O
,	O
</s>
<s>
enabling	O
the	O
LSTM	B-Algorithm
to	O
reset	O
its	O
own	O
state	O
.	O
</s>
<s>
Additionally	O
,	O
the	O
output	O
activation	B-Algorithm
function	I-Algorithm
was	O
omitted	O
.	O
</s>
<s>
2001	O
:	O
Gers	O
and	O
Schmidhuber	O
trained	O
LSTM	B-Algorithm
to	O
learn	O
languages	O
unlearnable	O
by	O
traditional	O
models	O
such	O
as	O
Hidden	O
Markov	O
Models	O
.	O
</s>
<s>
used	O
LSTM	B-Algorithm
for	O
meta-learning	B-General_Concept
(	O
i.e.	O
</s>
<s>
2004	O
:	O
First	O
successful	O
application	O
of	O
LSTM	B-Algorithm
to	O
speech	O
by	O
Schmidhuber	O
's	O
student	O
Alex	O
Graves	O
et	O
al	O
.	O
</s>
<s>
2005	O
:	O
First	O
publication	O
(	O
Graves	O
and	O
Schmidhuber	O
)	O
of	O
LSTM	B-Algorithm
with	O
full	O
backpropagation	B-Algorithm
through	I-Algorithm
time	I-Algorithm
and	O
of	O
bi-directional	O
LSTM	B-Algorithm
.	O
</s>
<s>
2005	O
:	O
Daan	O
Wierstra	O
,	O
Faustino	O
Gomez	O
,	O
and	O
Schmidhuber	O
trained	O
LSTM	B-Algorithm
by	O
neuroevolution	B-Algorithm
without	O
a	O
teacher	O
.	O
</s>
<s>
2006	O
:	O
Graves	O
,	O
Fernandez	O
,	O
Gomez	O
,	O
and	O
Schmidhuber	O
introduce	O
a	O
new	O
error	O
function	O
for	O
LSTM	B-Algorithm
:	O
Connectionist	B-Algorithm
Temporal	I-Algorithm
Classification	I-Algorithm
(	O
CTC	O
)	O
for	O
simultaneous	O
alignment	O
and	O
recognition	O
of	O
sequences	O
.	O
</s>
<s>
CTC-trained	O
LSTM	B-Algorithm
led	O
to	O
breakthroughs	O
in	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
trained	O
LSTM	B-Algorithm
to	O
control	O
robots	B-Algorithm
.	O
</s>
<s>
2007	O
:	O
Wierstra	O
,	O
Foerster	O
,	O
Peters	O
,	O
and	O
Schmidhuber	O
trained	O
LSTM	B-Algorithm
by	O
policy	O
gradients	O
for	O
reinforcement	O
learning	O
without	O
a	O
teacher	O
.	O
</s>
<s>
Hochreiter	O
,	O
Heuesel	O
,	O
and	O
Obermayr	O
applied	O
LSTM	B-Algorithm
to	O
protein	O
homology	O
detection	O
the	O
field	O
of	O
biology	O
.	O
</s>
<s>
2009	O
:	O
An	O
LSTM	B-Algorithm
trained	O
by	O
CTC	O
won	O
the	O
ICDAR	O
connected	O
handwriting	B-Application
recognition	I-Application
competition	O
.	O
</s>
<s>
introduced	O
neural	B-General_Concept
architecture	I-General_Concept
search	I-General_Concept
for	O
LSTM	B-Algorithm
.	O
</s>
<s>
2013	O
:	O
Alex	O
Graves	O
,	O
Abdel-rahman	O
Mohamed	O
,	O
and	O
Geoffrey	O
Hinton	O
used	O
LSTM	B-Algorithm
networks	O
as	O
a	O
major	O
component	O
of	O
a	O
network	O
that	O
achieved	O
a	O
record	O
17.7	O
%	O
phoneme	B-Language
error	O
rate	O
on	O
the	O
classic	O
TIMIT	B-Application
natural	O
speech	O
dataset	O
.	O
</s>
<s>
put	O
forward	O
a	O
simplified	O
variant	O
of	O
the	O
forget	O
gate	O
LSTM	B-Algorithm
called	O
Gated	B-Algorithm
recurrent	I-Algorithm
unit	I-Algorithm
(	O
GRU	O
)	O
.	O
</s>
<s>
2015	O
:	O
Google	O
started	O
using	O
an	O
LSTM	B-Algorithm
trained	O
by	O
CTC	O
for	O
speech	B-Application
recognition	I-Application
on	O
Google	O
Voice	O
.	O
</s>
<s>
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>
2016	O
:	O
Google	O
started	O
using	O
an	O
LSTM	B-Algorithm
to	O
suggest	O
messages	O
in	O
the	O
Allo	O
conversation	O
app	O
.	O
</s>
<s>
In	O
the	O
same	O
year	O
,	O
Google	O
released	O
the	O
Google	B-Algorithm
Neural	I-Algorithm
Machine	I-Algorithm
Translation	I-Algorithm
system	O
for	O
Google	O
Translate	O
which	O
used	O
LSTMs	B-Algorithm
to	O
reduce	O
translation	O
errors	O
by	O
60%	O
.	O
</s>
<s>
Apple	O
announced	O
in	O
its	O
Worldwide	O
Developers	O
Conference	O
that	O
it	O
would	O
start	O
using	O
the	O
LSTM	B-Algorithm
for	O
quicktype	O
in	O
the	O
iPhone	O
and	O
for	O
Siri	O
.	O
</s>
<s>
Amazon	O
released	O
Polly	B-Application
,	O
which	O
generates	O
the	O
voices	O
behind	O
Alexa	O
,	O
using	O
a	O
bidirectional	O
LSTM	B-Algorithm
for	O
the	O
text-to-speech	O
technology	O
.	O
</s>
<s>
2017	O
:	O
Facebook	O
performed	O
some	O
4.5	O
billion	O
automatic	B-Application
translations	I-Application
every	O
day	O
using	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
networks	O
.	O
</s>
<s>
Their	O
Time-Aware	O
LSTM	B-Algorithm
(	O
T-LSTM	O
)	O
performs	O
better	O
on	O
certain	O
data	O
sets	O
than	O
standard	O
LSTM	B-Algorithm
.	O
</s>
<s>
2018	O
:	O
OpenAI	O
used	O
LSTM	B-Algorithm
trained	O
by	O
policy	O
gradients	O
to	O
beat	O
humans	O
in	O
the	O
complex	O
video	O
game	O
of	O
Dota	O
2	O
,	O
and	O
to	O
control	O
a	O
human-like	O
robot	B-Algorithm
hand	O
that	O
manipulates	O
physical	O
objects	O
with	O
unprecedented	O
dexterity	O
.	O
</s>
<s>
2019	O
:	O
DeepMind	B-Application
used	O
LSTM	B-Algorithm
trained	O
by	O
policy	O
gradients	O
to	O
excel	O
at	O
the	O
complex	O
video	O
game	O
of	O
Starcraft	B-Application
II	I-Application
.	O
</s>
<s>
2021	O
:	O
According	O
to	O
Google	B-Library
Scholar	I-Library
,	O
in	O
2021	O
,	O
LSTM	B-Algorithm
was	O
cited	O
over	O
16,000	O
times	O
within	O
a	O
single	O
year	O
.	O
</s>
<s>
This	O
reflects	O
applications	O
of	O
LSTM	B-Algorithm
in	O
many	O
different	O
fields	O
including	O
healthcare	O
.	O
</s>
