14.02.2013 Views

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

118 Chapter 6. LNCS 3195:726-733, 2004<br />

PSfrag replacements<br />

(a) source images<br />

crosstalk<strong>in</strong>g error E1( Â, I)<br />

4<br />

3.5<br />

3<br />

2.5<br />

2<br />

1.5<br />

1<br />

0.5<br />

SOBI based on multi-dimensional autocovariances 5<br />

SOBI<br />

SOBI transposed images<br />

mdSOBI<br />

mdSOBI transposed images<br />

0<br />

0 10 20 30 40 50 60 70<br />

K<br />

(b) performance comparison<br />

Fig. 2. Comparison of SOBI and mdSOBI when applied to (unmixed) images from (a).<br />

The plot (b) plots the number K of time lags versus the crosstalk<strong>in</strong>g error E1 of the<br />

recovered matrix  and the unit matrix I; here  has been recovered by bot SOBI<br />

and mdSOBI given the images from (a) respectively the transposed images.<br />

after whiten<strong>in</strong>g of x(z1, . . . , zK). The jo<strong>in</strong>t diagonalizer then equals A except for<br />

permutation, given the generalized identifiability conditions from [4], theorem<br />

2. Therefore, also the identifiability result does not change, see [4]. In practice,<br />

we choose the (τ (k)<br />

1 , . . . , τ (k)<br />

M ) with <strong>in</strong>creas<strong>in</strong>g modulus for <strong>in</strong>creas<strong>in</strong>g k, but with<br />

the restriction τ (k)<br />

1 > 0 <strong>in</strong> order to avoid us<strong>in</strong>g the same autocovariances on the<br />

diagonal of the matrix twice.<br />

Often, data sets do not have any substantial long-distance autocorrelations,<br />

but quite high multi-dimensional close-distance correlations (see figure 1). When<br />

perform<strong>in</strong>g jo<strong>in</strong>t diagonalization, SOBI weighs each matrix equally strong, which<br />

can deteriorate the performance for large K, see simulation <strong>in</strong> section 4.<br />

Figure 2(a) shows an example, <strong>in</strong> which the images have considerable vertical<br />

structure, but rather random horizontal structure. Each of the two images<br />

consists of a concatenation of stripes of two images. For visual purposes, we<br />

chose the width of the stripes to be rather large with 16 pixels. Accord<strong>in</strong>g to<br />

the previous discussion we expect one-dimensional algorithms such as AMUSE<br />

and SOBI to perform well on the images, but badly (for number of time lags<br />

≫ 16) on the transposed images. If we apply AMUSE with τ = 20 to the images,<br />

we get excellent performance with a low crosstalk<strong>in</strong>g error with the unit<br />

matrix of 0.084; if we however apply AMUSE to the transposed images, the error<br />

is high with 1.1. This result is further confirmed by the comparison plot <strong>in</strong><br />

figure 2(b); mdSOBI performs equally well on the images and the transposed

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!