Mathematics in Independent Component Analysis
Mathematics in Independent Component Analysis
Mathematics in Independent Component Analysis
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
136 Chapter 8. Proc. NIPS 2006<br />
5.5<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
(a) S2<br />
6<br />
5.5<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
1.5<br />
7 8 9 10 11 12 13<br />
2<br />
14 3 4 5 6 7 8<br />
1.5<br />
9 3 3.5 4 4.5 5 5.5 6 6.5 7<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
−6.5 −6 −5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2<br />
(f) ( ˆ S1, ˆ S2)<br />
250<br />
200<br />
150<br />
100<br />
50<br />
(b) S3<br />
0<br />
−4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0<br />
(g) histogram of ˆ S3<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
14<br />
13<br />
12<br />
11<br />
10<br />
9<br />
2<br />
(c) S4<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0 0<br />
1<br />
(d) S5<br />
8<br />
−8<br />
−4.5 −4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 −7.5 −7 −6.5 −6 −5.5 −5 −4.5 −4 −3.5 −3 −2.5<br />
(h) S4<br />
−3<br />
−3.5<br />
−4<br />
−4.5<br />
−5<br />
−5.5<br />
−6<br />
−6.5<br />
−7<br />
−7.5<br />
(i) S5<br />
2<br />
3<br />
4<br />
1<br />
2<br />
3<br />
4<br />
5<br />
6<br />
7<br />
8<br />
5<br />
9<br />
10<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
−5<br />
0<br />
1 2 3 4 5 6 7 8 9 10<br />
(e) Â−1 A<br />
−1<br />
5<br />
−2<br />
4<br />
−3<br />
3<br />
2<br />
−4<br />
1<br />
−5 0<br />
(j) S6<br />
Figure 3: Example application of general ISA for unknown sizes m = (1, 2, 2, 2, 3). Shown are the<br />
scatter plots i.e. densities of the source components and the mix<strong>in</strong>g-separat<strong>in</strong>g map Â−1 A.<br />
also Cov(S) = I. Then I = Cov(X) = A Cov(S)A ⊤ = AA ⊤ so A is orthogonal. Due to the ISA<br />
assumptions, the fourth-order cross cumulants of the sources have to be trivial between different<br />
groups, and with<strong>in</strong> the Gaussians. In order to f<strong>in</strong>d transformations of the mixtures fulfill<strong>in</strong>g this<br />
property, we follow the idea of the JADE algorithmbut now <strong>in</strong> the ISA sett<strong>in</strong>g. We perform JBD of<br />
the (whitened) contracted quadricovariance matrices def<strong>in</strong>ed by Cij(X) := E � X ⊤ EijXXX ⊤� −<br />
Eij −E ⊤ ij −tr(Eij)I. Here RX := Cov(X) and Eij is a set of eigen-matrices of Cij, 1 ≤ i, j ≤ n.<br />
One simple choice is to use n 2 matrices Eij with zeros everywhere except 1 at <strong>in</strong>dex (i, j). More<br />
elaborate choices of eigen-matrices (with only n(n + 1)/2 or even n entries) are possible. The<br />
result<strong>in</strong>g algorithm, subspace-JADE (SJADE) not only performs NGCA by group<strong>in</strong>g Gaussians as<br />
one-dimensional components with trivial Cii’s, but also automatically f<strong>in</strong>ds the subspace partition<br />
m us<strong>in</strong>g the general JBD algorithm from section 2.3.<br />
4 Experimental results<br />
In a first example, we consider a general ISA problem <strong>in</strong> dimension n = 10 with the unknown<br />
partition m = (1, 2, 2, 2, 3). In order to generate 2- and 3-dimensional irreducible random vectors,<br />
we decided to follow the nice visual ideas from [10] and to draw samples from a density follow<strong>in</strong>g<br />
a known shape — <strong>in</strong> our case 2d-letters or 3d-geometrical shapes. The chosen sources densities are<br />
shown <strong>in</strong> figure 3(a-d). Another 1-dimensional source follow<strong>in</strong>g a uniform distribution was constructed.<br />
Altogether 104 samples were used. The sources S were mixed by a mix<strong>in</strong>g matrix A with<br />
coefficients uniformly randomly sampled from [−1, 1] to give mixtures X = AS. The recovered<br />
mix<strong>in</strong>g matrix  was then estimated us<strong>in</strong>g the above block-JADE algorithm with unknown block<br />
size; we observed that the method is quite sensitive to the choice of the threshold (here θ = 0.015).<br />
Figure 3(e) shows the composed mix<strong>in</strong>g-separat<strong>in</strong>g system Â−1A; clearly the matrices are equal<br />
except for block permutation and scal<strong>in</strong>g, which experimentally confirms theorem 1.8. The algorithm<br />
found a partition ˆm = (1, 1, 1, 2, 2, 3), so one 2d-source was mis<strong>in</strong>terpreted as two 1d-sources,<br />
but by us<strong>in</strong>g previous knowledge comb<strong>in</strong>ation of the correct two 1d-sources yields the orig<strong>in</strong>al 2dsource.<br />
The result<strong>in</strong>g recovered sources ˆ S := Â−1X, figures 3(f-j), then equal the orig<strong>in</strong>al sources<br />
except for permutation and scal<strong>in</strong>g with<strong>in</strong> the sources — which <strong>in</strong> the higher-dimensional cases<br />
implies transformations such as rotation of the underly<strong>in</strong>g images or shapes. When apply<strong>in</strong>g ICA<br />
(1-ISA) to the above mixtures, we cannot expect to recover the orig<strong>in</strong>al sources as expla<strong>in</strong>ed <strong>in</strong><br />
figure 1; however, some algorithms might recover the sources up to permutation. Indeed, SJADE<br />
equals JADE with additional permutation recovery because the jo<strong>in</strong>t block diagonalization is per-