<s>
A	O
relation	B-Algorithm
network	I-Algorithm
(	O
RN	O
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
component	O
with	O
a	O
structure	O
that	O
can	O
reason	O
about	O
relations	O
among	O
objects	O
.	O
</s>
<s>
In	O
June	O
2017	O
,	O
DeepMind	B-Application
announced	O
the	O
first	O
relation	B-Algorithm
network	I-Algorithm
.	O
</s>
<s>
It	O
claimed	O
that	O
the	O
technology	O
had	O
achieved	O
"	O
superhuman	O
"	O
performance	O
on	O
multiple	O
question-answering	B-Algorithm
problem	O
sets	O
.	O
</s>
<s>
"	O
)	O
CleVR	B-Algorithm
(	O
state	O
description	O
)	O
96.4	O
%Images	O
represented	O
by	O
state	O
description	O
matrices	O
.	O
</s>
<s>
RNs	O
constrain	O
the	O
functional	O
form	O
of	O
a	O
neural	B-Architecture
network	I-Architecture
to	O
capture	O
the	O
common	O
properties	O
of	O
relational	O
reasoning	O
.	O
</s>
<s>
These	O
properties	O
are	O
explicitly	O
added	O
to	O
the	O
system	O
,	O
rather	O
than	O
established	O
by	O
learning	O
just	O
as	O
the	O
capacity	O
to	O
reason	O
about	O
spatial	O
,	O
translation-invariant	O
properties	O
is	O
explicitly	O
part	O
of	O
convolutional	B-Architecture
neural	I-Architecture
networks	I-Architecture
(	O
CNN	B-Architecture
)	O
.	O
</s>
<s>
fφ	O
and	O
gθ	O
are	O
multilayer	B-Algorithm
perceptrons	I-Algorithm
,	O
while	O
the	O
2	O
parameters	O
are	O
learnable	O
synaptic	O
weights	O
.	O
</s>
<s>
Image	O
(	O
128x128	O
pixel	O
)	O
processing	O
is	O
done	O
with	O
a	O
4-layer	O
CNN	B-Architecture
.	O
</s>
<s>
Outputs	O
from	O
the	O
CNN	B-Architecture
are	O
treated	O
as	O
the	O
objects	O
for	O
relation	O
analysis	O
,	O
without	O
regard	O
for	O
what	O
those	O
"	O
objects	O
"	O
explicitly	O
represent	O
.	O
</s>
<s>
Questions	O
were	O
processed	O
with	O
a	O
long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
network	O
.	O
</s>
