<s>
In	O
artificial	B-Application
intelligence	I-Application
,	O
a	O
differentiable	B-Algorithm
neural	I-Algorithm
computer	I-Algorithm
(	O
DNC	O
)	O
is	O
a	O
memory	O
augmented	O
neural	B-Architecture
network	I-Architecture
architecture	O
(	O
MANN	O
)	O
,	O
which	O
is	O
typically	O
(	O
but	O
not	O
by	O
definition	O
)	O
recurrent	O
in	O
its	O
implementation	O
.	O
</s>
<s>
of	O
DeepMind	B-Application
.	O
</s>
<s>
DNC	O
indirectly	O
takes	O
inspiration	O
from	O
Von-Neumann	B-Architecture
architecture	I-Architecture
,	O
making	O
it	O
likely	O
to	O
outperform	O
conventional	O
architectures	O
in	O
tasks	O
that	O
are	O
fundamentally	O
algorithmic	O
that	O
cannot	O
be	O
learned	O
by	O
finding	O
a	O
decision	B-General_Concept
boundary	I-General_Concept
.	O
</s>
<s>
This	O
attention	O
span	O
allows	O
the	O
user	O
to	O
feed	O
complex	O
data	B-General_Concept
structures	I-General_Concept
such	O
as	O
graphs	B-Application
sequentially	O
,	O
and	O
recall	O
them	O
for	O
later	O
use	O
.	O
</s>
<s>
Furthermore	O
,	O
they	O
can	O
learn	O
aspects	O
of	O
symbolic	B-Algorithm
reasoning	I-Algorithm
and	O
apply	O
it	O
to	O
working	O
memory	O
.	O
</s>
<s>
A	O
neural	B-Architecture
network	I-Architecture
without	O
memory	O
would	O
typically	O
have	O
to	O
learn	O
about	O
each	O
transit	O
system	O
from	O
scratch	O
.	O
</s>
<s>
On	O
graph	O
traversal	O
and	O
sequence-processing	O
tasks	O
with	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
DNCs	O
performed	O
better	O
than	O
alternatives	O
such	O
as	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
or	O
a	O
neural	B-Algorithm
turing	I-Algorithm
machine	I-Algorithm
.	O
</s>
<s>
With	O
a	O
reinforcement	O
learning	O
approach	O
to	O
a	O
block	O
puzzle	O
problem	O
inspired	O
by	O
SHRDLU	O
,	O
DNC	O
was	O
trained	O
via	O
curriculum	O
learning	O
,	O
and	O
learned	O
to	O
make	O
a	O
plan	B-Application
.	O
</s>
<s>
It	O
performed	O
better	O
than	O
a	O
traditional	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
.	O
</s>
<s>
DNC	O
networks	O
were	O
introduced	O
as	O
an	O
extension	O
of	O
the	O
Neural	B-Algorithm
Turing	I-Algorithm
Machine	I-Algorithm
(	O
NTM	O
)	O
,	O
with	O
the	O
addition	O
of	O
memory	O
attention	O
mechanisms	O
that	O
control	O
where	O
the	O
memory	O
is	O
stored	O
,	O
and	O
temporal	O
attention	O
that	O
records	O
the	O
order	O
of	O
events	O
.	O
</s>
<s>
This	O
structure	O
allows	O
DNCs	O
to	O
be	O
more	O
robust	O
and	O
abstract	O
than	O
a	O
NTM	O
,	O
and	O
still	O
perform	O
tasks	O
that	O
have	O
longer-term	O
dependencies	O
than	O
some	O
predecessors	O
such	O
as	O
Long	B-Algorithm
Short	I-Algorithm
Term	I-Algorithm
Memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
.	O
</s>
<s>
This	O
makes	O
it	O
possible	O
to	O
optimize	O
them	O
efficiently	O
using	O
gradient	B-Algorithm
descent	I-Algorithm
.	O
</s>
<s>
The	O
DNC	O
model	O
is	O
similar	O
to	O
the	O
Von	B-Architecture
Neumann	I-Architecture
architecture	I-Architecture
,	O
and	O
because	O
of	O
the	O
resizability	O
of	O
memory	O
,	O
it	O
is	O
Turing	B-Algorithm
complete	I-Algorithm
.	O
</s>
<s>
This	O
can	O
be	O
achieved	O
by	O
using	O
an	O
approximate	O
nearest	O
neighbor	O
algorithm	O
,	O
such	O
as	O
Locality-sensitive	B-Algorithm
hashing	I-Algorithm
,	O
or	O
a	O
random	O
k-d	B-Data_Structure
tree	I-Data_Structure
like	O
Fast	O
Library	O
for	O
Approximate	O
Nearest	O
Neighbors	O
from	O
UBC	O
.	O
</s>
<s>
Training	O
using	O
synthetic	O
gradients	O
performs	O
considerably	O
better	O
than	O
Backpropagation	B-Algorithm
through	I-Algorithm
time	I-Algorithm
(	O
BPTT	B-Algorithm
)	O
.	O
</s>
