Digital Signal Processing Reference
In-Depth Information
5
0
5
0
20
40
60
80
100
1
2
0
5
0
5
0
20
40
60
80
100
1
2
0
5
0
5
0
20
40
60
80
100
1
2
0
5
0
5
0
20
40
60
80
100
1
2
0
2
0
2
0
20
40
60
80
100
1
2
0
2
0
2
0
20
40
60
80
100
120
Figure 13.6
The five independent components of the data set calculated using FastICA with
deflation and pow3-nonlinearity after whitening and PCA dimension reduction to
five dimensions. Below the five components, the cell/non cell functions (
−
1or1)of
the samples are plotted for comparison. The crosscorrelations of each signal with
these functions are
−
0
.
055, 0
.
11, 0
.
37,
−
0
.
081, and 0
.
90. Visual comparison already
confirms good correspondence of the fifth IC with the cell/non cell function.
data separation, albeit by a rather small factor in this case.
Neural network learning
In the previous text, we saw that even unsupervised methods were
sucient to detect an acceptable cell classifier in the given BrdU-labeled
cell experiment. However, performance is somewhat weak, because the
knowledge of cell/non cell labels was exploited only afterward. Also, if
more complicated structures are to be analyzed, nonlinear classifiers turn
out to be preferable. In order to allow for training as well as flexibility
regarding the strength of possible nonlinearities, we will use neural
networks to learn such a cell classifier. Depending on the applications
Search WWH ::
Custom Search