<s>
A	O
Siamese	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
(	O
sometimes	O
called	O
a	O
twin	O
neural	B-Architecture
network	I-Architecture
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
that	O
uses	O
the	O
same	O
weights	O
while	O
working	O
in	O
tandem	O
on	O
two	O
different	O
input	O
vectors	O
to	O
compute	O
comparable	O
output	O
vectors	O
.	O
</s>
<s>
This	O
is	O
similar	O
to	O
comparing	O
fingerprints	B-Algorithm
but	O
can	O
be	O
described	O
more	O
technically	O
as	O
a	O
distance	O
function	O
for	O
locality-sensitive	B-Algorithm
hashing	I-Algorithm
.	O
</s>
<s>
It	O
is	O
possible	O
to	O
build	O
an	O
architecture	O
that	O
is	O
functionally	O
similar	O
to	O
a	O
siamese	B-Algorithm
network	I-Algorithm
but	O
implements	O
a	O
slightly	O
different	O
function	O
.	O
</s>
<s>
Uses	O
of	O
similarity	O
measures	O
where	O
a	O
twin	O
network	O
might	O
be	O
used	O
are	O
such	O
things	O
as	O
recognizing	B-Application
handwritten	I-Application
checks	O
,	O
automatic	O
detection	B-General_Concept
of	I-General_Concept
faces	I-General_Concept
in	O
camera	O
images	O
,	O
and	O
matching	O
queries	O
with	O
indexed	O
documents	O
.	O
</s>
<s>
Learning	O
in	O
twin	O
networks	O
can	O
be	O
done	O
with	O
triplet	B-Algorithm
loss	I-Algorithm
or	O
contrastive	O
loss	O
.	O
</s>
<s>
For	O
learning	O
by	O
triplet	B-Algorithm
loss	I-Algorithm
a	O
baseline	O
vector	O
(	O
anchor	O
image	O
)	O
is	O
compared	O
against	O
a	O
positive	O
vector	O
(	O
truthy	O
image	O
)	O
and	O
a	O
negative	O
vector	O
(	O
falsy	O
image	O
)	O
.	O
</s>
<s>
In	O
particular	O
,	O
the	O
triplet	B-Algorithm
loss	I-Algorithm
algorithm	O
is	O
often	O
defined	O
with	O
squared	O
Euclidean	O
(	O
which	O
unlike	O
Euclidean	O
,	O
does	O
not	O
have	O
triangle	O
inequality	O
)	O
distance	O
at	O
its	O
core	O
.	O
</s>
<s>
This	O
can	O
be	O
further	O
subdivided	O
in	O
at	O
least	O
Unsupervised	B-General_Concept
learning	I-General_Concept
and	O
Supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
After	O
being	O
first	O
introduced	O
in	O
2016	O
,	O
Twin	O
fully	O
convolutional	O
network	O
has	O
been	O
used	O
in	O
many	O
High-performance	O
Real-time	O
Object	O
Tracking	O
Neural	B-Architecture
Networks	I-Architecture
.	O
</s>
