Digital Signal Processing Reference
In-Depth Information
1. Serial estimation: In this approach, the components are sequen-
tially estimated. This means that all extracting vectors
w
i
must
be, necessarily, orthogonal to the previously obtained vectors. For
this purpose, one can employ the Gram-Schmidt orthogonalization
method [128]. This serial approach is also known in the literature as
the deflation approach [89].
2. Parallel estimation: In this case, a certain number of sources will
be estimated at once, adapting in parallel the vectors
w
i
. However,
since it is required that all vectors be orthogonal, an additional
step of orthonormalization is required, and the Gram-Schmidt
procedure can be employed again [148].
6.2.2.5 The Infomax Principle and the Maximum Likelihood Approach
Another interesting approach to perform ICA is the so-called Infomax prin-
ciple, introduced in the context of BSS by Bell and Sejnowski [31], even
though key results had already been established in a different context [212].
The approach is based on the concepts issued from the field of neural net-
works. Neural networks will be discussed in more detail in Chapter 7, but
for the moment, it suffices to consider that one possible structure of a neural
network is composed of a linear portion and a set of nonlinearities.
Let us consider the structure depicted in Figure 6.4, where
A
represents
the mixing system. The separating system is an artificial neural network com-
posed by a linear part (the matrix
W
) and a set of nonlinearities
f
i
(
·
)
, each one
applied to a particular output
y
i
, so that we define the vector.
f
1
y
1
f
2
y
2
f
N
y
N
T
f
y
=
···
(6.37)
The nonlinear functions
f
i
(
·
)
are monotonically increasing, with
f
i
(
−∞
)
=
1.
According to the Infomax principle, the coefficients of the neural net-
work should be adjusted in order to maximize the amount of information
that flows from the inputs to the outputs, which means that
W
should be
chosen to maximize the mutual information between
x
and
z
, thus leading to
0and
f
i
(
∞
)
=
s
1
(
n
)
x
1
(
n
)
y
1
(
n
)
f
1
(
.
)
z
1
(
n
)
W
A
s
N
(
n
)
x
N
(
n
)
y
N
(
n
)
f
N
(
.
)
z
N
(
n
)
FIGURE 6.4
Structure of an artificial neural network.