<s>
A	O
residual	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
(	O
ResNet	O
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
(	O
ANN	O
)	O
.	O
</s>
<s>
It	O
is	O
a	O
gateless	O
or	O
open-gated	O
variant	O
of	O
the	O
HighwayNet	B-General_Concept
,	O
the	O
first	O
working	O
very	O
deep	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
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
Skip	B-Algorithm
connections	I-Algorithm
or	O
shortcuts	O
are	O
used	O
to	O
jump	O
over	O
some	O
layers	O
(	O
HighwayNets	B-General_Concept
may	O
also	O
learn	O
the	O
skip	O
weights	O
themselves	O
through	O
an	O
additional	O
weight	O
matrix	O
for	O
their	O
gates	O
)	O
.	O
</s>
<s>
Typical	O
ResNet	O
models	O
are	O
implemented	O
with	O
double	O
-	O
or	O
triple	O
-	O
layer	O
skips	O
that	O
contain	O
nonlinearities	O
(	O
ReLU	B-Algorithm
)	O
and	O
batch	B-General_Concept
normalization	I-General_Concept
in	O
between	O
.	O
</s>
<s>
Models	O
with	O
several	O
parallel	O
skips	O
are	O
referred	O
to	O
as	O
DenseNets	B-Algorithm
.	O
</s>
<s>
In	O
the	O
context	O
of	O
residual	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
,	O
a	O
non-residual	O
network	O
may	O
be	O
described	O
as	O
a	O
plain	O
network	O
.	O
</s>
<s>
If	O
not	O
,	O
then	O
an	O
explicit	O
weight	O
matrix	O
should	O
be	O
learned	O
for	O
the	O
skipped	O
connection	O
(	O
a	O
HighwayNet	B-General_Concept
should	O
be	O
used	O
)	O
.	O
</s>
<s>
This	O
speeds	O
learning	O
by	O
reducing	O
the	O
impact	O
of	O
vanishing	B-Algorithm
gradients	I-Algorithm
,	O
as	O
there	O
are	O
fewer	O
layers	O
to	O
propagate	O
through	O
.	O
</s>
<s>
The	O
network	O
then	O
gradually	O
restores	O
the	O
skipped	O
layers	O
as	O
it	O
learns	O
the	O
feature	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
A	O
neural	B-Architecture
network	I-Architecture
without	O
residual	O
parts	O
explores	O
more	O
of	O
the	O
feature	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
A	O
residual	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
was	O
used	O
to	O
win	O
the	O
ImageNet	B-General_Concept
2015	O
competition	O
,	O
and	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>
the	O
activation	B-Algorithm
function	I-Algorithm
for	O
layer	O
,	O
</s>
<s>
If	O
the	O
number	O
of	O
vertices	O
on	O
layer	O
equals	O
the	O
number	O
of	O
vertices	O
on	O
layer	O
and	O
if	O
is	O
the	O
identity	O
matrix	O
,	O
then	O
forward	O
propagation	O
through	O
the	O
activation	B-Algorithm
function	I-Algorithm
simplifies	O
to	O
In	O
this	O
case	O
,	O
the	O
connection	O
between	O
layers	O
and	O
is	O
called	O
an	O
identity	O
block	O
.	O
</s>
<s>
a	O
learning	B-General_Concept
rate	I-General_Concept
(	O
,	O
</s>
<s>
If	O
they	O
can	O
be	O
updated	O
,	O
the	O
rule	O
is	O
an	O
ordinary	O
backpropagation	B-Algorithm
update	O
rule	O
.	O
</s>
