<s>
A	O
neural	B-Algorithm
Turing	I-Algorithm
machine	I-Algorithm
(	O
NTM	O
)	O
is	O
a	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
model	O
of	O
a	O
Turing	B-Architecture
machine	I-Architecture
.	O
</s>
<s>
NTMs	O
combine	O
the	O
fuzzy	O
pattern	B-Language
matching	I-Language
capabilities	O
of	O
neural	B-Architecture
networks	I-Architecture
with	O
the	O
algorithmic	O
power	O
of	O
programmable	O
computers	O
.	O
</s>
<s>
An	O
NTM	O
has	O
a	O
neural	B-Architecture
network	I-Architecture
controller	O
coupled	O
to	O
external	O
memory	O
resources	O
,	O
which	O
it	O
interacts	O
with	O
through	O
attentional	O
mechanisms	O
.	O
</s>
<s>
The	O
memory	O
interactions	O
are	O
differentiable	O
end-to-end	O
,	O
making	O
it	O
possible	O
to	O
optimize	O
them	O
using	O
gradient	B-Algorithm
descent	I-Algorithm
.	O
</s>
<s>
An	O
NTM	O
with	O
a	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
network	O
controller	O
can	O
infer	O
simple	O
algorithms	O
such	O
as	O
copying	O
,	O
sorting	O
,	O
and	O
associative	O
recall	O
from	O
examples	O
alone	O
.	O
</s>
<s>
The	O
first	O
stable	O
open-source	O
implementation	O
was	O
published	O
in	O
2018	O
at	O
the	O
27th	O
International	O
Conference	O
on	O
Artificial	O
Neural	B-Architecture
Networks	I-Architecture
,	O
receiving	O
a	O
best-paper	O
award	O
.	O
</s>
<s>
The	O
developers	O
either	O
report	O
that	O
the	O
gradients	B-Algorithm
of	O
their	O
implementation	O
sometimes	O
become	O
NaN	O
during	O
training	O
for	O
unknown	O
reasons	O
and	O
cause	O
training	O
to	O
fail	O
;	O
report	O
slow	O
convergence	O
;	O
or	O
do	O
not	O
report	O
the	O
speed	O
of	O
learning	O
of	O
their	O
implementation	O
.	O
</s>
<s>
Differentiable	B-Algorithm
neural	I-Algorithm
computers	I-Algorithm
are	O
an	O
outgrowth	O
of	O
Neural	B-Algorithm
Turing	I-Algorithm
machines	I-Algorithm
,	O
with	O
attention	B-General_Concept
mechanisms	I-General_Concept
that	O
control	O
where	O
the	O
memory	O
is	O
active	O
,	O
and	O
improve	O
performance	O
.	O
</s>
