Digital Signal Processing Reference
In-Depth Information
1.3
1.2
1.1
1
0.9
0.8
0.7
0.6
0.5
0
20
40
60
80
100
120
140
160
180
200
Figure 4.8
Kurtosis maximization: absolute kurtosis versus angle. The function
α →|
kurt((cos(
α
)sin(
α
))
z
)
|
is plotted with the uniform
z
from figure 4.4.
We have seen that prewhitening (i.e. PCA) is essential for this
algorithm — it reduces the search dimension by making the problem
easily accessible. The above equation can be interpreted as finding the
projection onto the line given by
w
such that
z
along this line is maximal
non Gaussian.
In figures 4.8 and 4.9, the absolute kurtosis is plotted for the uniform-
source example respectively the Laplacian example from above.
Gradient ascent kurtosis maximization
In practice local algorithms are often interesting. A differentiable func-
tion
f
:
n
can be maximized by local updates in the direction of
its gradient (which points to the direction of greatest ascent). Given a
suciently small
learning rate η>
0andastartingpoint
x
(0)
R
→ R
n
,
∈ R
local maxima of
f
can be found by iterating
x
(
t
+1)=
x
(
t
)+
η
Δ
x
(
t
)
with
f
(
x
(
t
)) =
∂f
Δ
x
(
t
)=(
Df
)(
x
(
t
))
=
∂
x
(
x
(
t
))
being the gradient of
f
at
x
(
t
). This algorithm is called
gradient ascent
.
Often, the learning rate
η
is chosen to be dependent on the time
t
,and
∇
Search WWH ::
Custom Search