14.02.2013 Views

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

274 Chapter 20. Signal Process<strong>in</strong>g 86(3):603-623, 2006<br />

2.3 Measures used for comparison<br />

In order to compare the recovered signals with the artificial sources, we simply<br />

compare the mix<strong>in</strong>g matrices. For this we employ Amari’s separation performance<br />

<strong>in</strong>dex [41], which is given by the equation<br />

⎛<br />

⎞<br />

n� n� |pij|<br />

n�<br />

�<br />

n�<br />

�<br />

|pij|<br />

E1(P) = ⎝<br />

− 1⎠<br />

+<br />

− 1<br />

i=1 j=1 maxk |pik|<br />

j=1 i=1 maxk |pkj|<br />

where P = (pij) = Â+ A, A be<strong>in</strong>g the real mix<strong>in</strong>g matrix and Â+ the<br />

pseudo<strong>in</strong>verse of its estimation Â. Note that E1(P) ≤ 2n(n − 1). For both<br />

the artificial and the real signals, we calculate E1( Â+ 1 Â2), where Âi are the<br />

two recovered mix<strong>in</strong>g matrices.<br />

Furthermore, we also want to study the recovered signals. So <strong>in</strong> order to be able<br />

to compare between different methods, to each pair of components obta<strong>in</strong>ed<br />

with each method, as well as to the source signals and the synthetic s-EMG<br />

channels, we apply the follow<strong>in</strong>g equivalence measures: Pr<strong>in</strong>cipe’s quadratic<br />

mutual <strong>in</strong>formation (QMI) [42], Kullback-Leibler <strong>in</strong>formation distance (K-LD)<br />

[43], Renyi’s entropy measure [43], mutual <strong>in</strong>formation measure (MuIn) [42],<br />

Rosenblatt’s squared distance functional (RSD) [43], Skaug and Tjøstheim’s<br />

weighted difference (STW) [43] and cross-correlation (Xcor). All the measures<br />

are normalized with respect to the maximum value obta<strong>in</strong>ed when applied<br />

to each component with itself; that is, we divide by the maximum of each<br />

comparison matrix diagonal (maximum mutual <strong>in</strong>formation).<br />

For the calculation of the above-mentioned <strong>in</strong>dices it is necessary to estimate<br />

both jo<strong>in</strong>t and marg<strong>in</strong>al probability density function of the signals. We have<br />

decide to use the data-driven method Kn-nearest-neighbour (KnNN) [44] (for<br />

details, refer to [32]). Furthermore, <strong>in</strong> order to compare separation performance<br />

<strong>in</strong> the presence of noise, we measure the strength of a one-dimensional<br />

signal S versus additive noise ˜ S us<strong>in</strong>g the signal-to-noise ratio (SNR) def<strong>in</strong>ed<br />

by<br />

SNR(S, ˜ �S�<br />

S) := 20 log10 �S − ˜ S� .<br />

3 Results<br />

In this section we compare the various models for source separation when<br />

applied to both toy and real s-EMG record<strong>in</strong>gs.<br />

13

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

Saved successfully!

Ooh no, something went wrong!