<s>
FastICA	B-Algorithm
is	O
an	O
efficient	O
and	O
popular	O
algorithm	O
for	O
independent	B-Algorithm
component	I-Algorithm
analysis	I-Algorithm
invented	O
by	O
Aapo	O
Hyvärinen	O
at	O
Helsinki	O
University	O
of	O
Technology	O
.	O
</s>
<s>
Like	O
most	O
ICA	O
algorithms	O
,	O
FastICA	B-Algorithm
seeks	O
an	O
orthogonal	O
rotation	O
of	O
prewhitened	O
data	O
,	O
through	O
a	O
fixed-point	O
iteration	B-Algorithm
scheme	I-Algorithm
,	O
that	O
maximizes	O
a	O
measure	O
of	O
non-Gaussianity	O
of	O
the	O
rotated	O
components	O
.	O
</s>
<s>
FastICA	B-Algorithm
can	O
also	O
be	O
alternatively	O
derived	O
as	O
an	O
approximative	O
Newton	O
iteration	O
.	O
</s>
<s>
The	O
input	O
data	O
matrix	O
must	O
be	O
prewhitened	O
,	O
or	O
centered	O
and	O
whitened	O
,	O
before	O
applying	O
the	O
FastICA	B-Algorithm
algorithm	O
to	O
it	O
.	O
</s>
<s>
Whitening	O
the	O
data	O
requires	O
a	O
linear	B-Architecture
transformation	I-Architecture
of	O
the	O
centered	O
data	O
so	O
that	O
the	O
components	O
of	O
are	O
uncorrelated	O
and	O
have	O
variance	O
one	O
.	O
</s>
<s>
To	O
measure	O
non-Gaussianity	O
,	O
FastICA	B-Algorithm
relies	O
on	O
a	O
nonquadratic	O
nonlinear	O
function	O
,	O
its	O
first	O
derivative	O
,	O
and	O
its	O
second	O
derivative	O
.	O
</s>
<s>
The	O
steps	O
for	O
extracting	O
the	O
weight	O
vector	O
for	O
single	O
component	O
in	O
FastICA	B-Algorithm
are	O
the	O
following	O
:	O
</s>
<s>
The	O
single	O
unit	O
iterative	B-Algorithm
algorithm	I-Algorithm
estimates	O
only	O
one	O
weight	O
vector	O
which	O
extracts	O
a	O
single	O
component	O
.	O
</s>
