<s>
The	O
swish	B-Algorithm
function	I-Algorithm
is	O
a	O
mathematical	O
function	O
defined	O
as	O
follows	O
:	O
</s>
<s>
For	O
β	O
=	O
1	O
,	O
the	O
function	O
becomes	O
equivalent	O
to	O
the	O
Sigmoid	B-Algorithm
Linear	O
Unit	O
or	O
SiLU	O
,	O
first	O
proposed	O
alongside	O
the	O
GELU	B-Algorithm
in	O
2016	O
.	O
</s>
<s>
The	O
SiLU	O
was	O
later	O
rediscovered	O
in	O
2017	O
as	O
the	O
Sigmoid-weighted	B-Algorithm
Linear	I-Algorithm
Unit	I-Algorithm
(	O
SiL	O
)	O
function	O
used	O
in	O
reinforcement	O
learning	O
.	O
</s>
<s>
The	O
SiLU/SiL	O
was	O
then	O
rediscovered	O
as	O
the	O
swish	B-Algorithm
over	O
a	O
year	O
after	O
its	O
initial	O
discovery	O
,	O
originally	O
proposed	O
without	O
the	O
learnable	O
parameter	O
β	O
,	O
so	O
that	O
β	O
implicitly	O
equalled	O
1	O
.	O
</s>
<s>
The	O
swish	B-Algorithm
paper	O
was	O
then	O
updated	O
to	O
propose	O
the	O
activation	O
with	O
the	O
learnable	O
parameter	O
β	O
,	O
though	O
researchers	O
usually	O
let	O
β	O
=	O
1	O
and	O
do	O
not	O
use	O
the	O
learnable	O
parameter	O
β	O
.	O
</s>
<s>
With	O
β	O
→	O
∞	O
,	O
the	O
sigmoid	B-Algorithm
component	O
approaches	O
a	O
0-1	O
function	O
pointwise	O
,	O
so	O
swish	B-Algorithm
approaches	O
the	O
ReLU	B-Algorithm
function	O
pointwise	O
.	O
</s>
<s>
Thus	O
,	O
it	O
can	O
be	O
viewed	O
as	O
a	O
smoothing	O
function	O
which	O
nonlinearly	O
interpolates	O
between	O
a	O
linear	O
function	O
and	O
the	O
ReLU	B-Algorithm
function	O
.	O
</s>
<s>
This	O
function	O
uses	O
non-monotonicity	O
,	O
and	O
may	O
have	O
influenced	O
the	O
proposal	O
of	O
other	O
activation	B-Algorithm
functions	I-Algorithm
with	O
this	O
property	O
such	O
as	O
Mish	O
.	O
</s>
<s>
When	O
considering	O
positive	O
values	O
,	O
Swish	B-Algorithm
is	O
a	O
particular	O
case	O
of	O
sigmoid	B-Algorithm
shrinkage	O
function	O
defined	O
in	O
(	O
see	O
the	O
doubly	O
parameterized	O
sigmoid	B-Algorithm
shrinkage	O
form	O
given	O
by	O
Equation	O
(	O
3	O
)	O
of	O
this	O
reference	O
)	O
.	O
</s>
<s>
In	O
2017	O
,	O
after	O
performing	O
analysis	O
on	O
ImageNet	B-General_Concept
data	O
,	O
researchers	O
from	O
Google	B-Application
indicated	O
that	O
using	O
this	O
function	O
as	O
an	O
activation	B-Algorithm
function	I-Algorithm
in	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
improves	O
the	O
performance	O
,	O
compared	O
to	O
ReLU	B-Algorithm
and	O
sigmoid	B-Algorithm
functions	I-Algorithm
.	O
</s>
<s>
It	O
is	O
believed	O
that	O
one	O
reason	O
for	O
the	O
improvement	O
is	O
that	O
the	O
swish	B-Algorithm
function	I-Algorithm
helps	O
alleviate	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
during	O
backpropagation	B-Algorithm
.	O
</s>
