Learning Mixtures of Independent Component Analysers - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 68

Fig. 3.1 First simulation
experiment—BSS in ICA
mixture datasets

50

45

40

35

Mixca
FastIca
JADE
InfoMax
Extended InfoMax
Radical
Npica
Kernel-ICA
TDSEP

30

25

20

15

10

5

0 500 1000 1500 2000

N

Table 3.4 Sources used in the simulation experiment

Source # Source type Skewness Kurtosis pdf plot

1 Uniform 0.0 -1.2

2 Laplacian (b ¼ 1) 0.0 3

3 Normal 0.0 0.0

4 Rayleigh (b ¼ 1) 0.631 0.245

5 K-noise (m ¼ 1) 1.28 5.38

performance in BSS, and the Mixca algorithm was configured to estimate the
parameters for only one ICA. TDSEP, which is an algorithm based on exploiting
signal time structure, was not tested in the second experiment since this experi-
ment included i.i.d. sources. If the sources involved in the problem have a time
structure, the second order based methods as TDSEP are appropriate. In this case,
the limitation is not due to the Gaussianity of the signals; it is due to the correlation
functions. These functions can extract Gaussian signals, but with different spectra
by diagonalizing correlation matrices for different time lags. Second order based
methods are different from ICA traditional methods, which recover non Gaussian