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.
Chapter 19. LNCS 3195:977-984, 2004 253<br />
978 Ingo R. Keck et al.<br />
In this text we compare these analysis techniques <strong>in</strong> a study of an auditory<br />
task. We show an example where traditional model based methods do not yield<br />
reasonable results. Rather bl<strong>in</strong>d source separation techniques have to be used<br />
to get mean<strong>in</strong>gful and <strong>in</strong>terest<strong>in</strong>g results concern<strong>in</strong>g the networks of activations<br />
related to a comb<strong>in</strong>ed word recognition and motor task.<br />
1.1 Model Based Approach: General L<strong>in</strong>ear Model<br />
The general l<strong>in</strong>ear model as a k<strong>in</strong>d of regression analysis has been the classic<br />
way to analyze fMRI data <strong>in</strong> the past [3]. Basically it uses second order statistics<br />
to f<strong>in</strong>d the voxels whose activations correlate best to given time courses. The<br />
measured signal for each voxel <strong>in</strong> time y =(y(t1), ..., y(tn)) T is written as a<br />
l<strong>in</strong>ear comb<strong>in</strong>ation of <strong>in</strong>dependent variables y = Xb + e, with the vector b of<br />
regression coefficients and the matrix X of the <strong>in</strong>dependent variables which <strong>in</strong><br />
case of an fMRI-analysis consist of the assumed time courses <strong>in</strong> the data and<br />
additional filters to account for the serial correlation of fMRI data. The residual<br />
error e ought to be m<strong>in</strong>imized. The normal equation X T Xb = X T y of the<br />
problem is solved by b =(X T X) −1 X T y and has a unique solution if XX T has<br />
full rank. F<strong>in</strong>ally a significance test us<strong>in</strong>g e is applied to estimate the statistical<br />
significance of the found correlation.<br />
As the model X must be known <strong>in</strong> advance to calculate b, this method is<br />
called “model-based”. It can be used to test the accuracy of a given model, but<br />
cannot by itself f<strong>in</strong>d a better suited model even if one exists.<br />
1.2 Model Free Approach:<br />
BSS Us<strong>in</strong>g <strong>Independent</strong> <strong>Component</strong> <strong>Analysis</strong><br />
In case of fMRI data bl<strong>in</strong>d source separation refers to the problem of separat<strong>in</strong>g<br />
a given sensor signal, i.e. the fMRI data at the time t<br />
x(t) =A [s(t)+snoise(t)] =<br />
n�<br />
aisi(t)+<br />
i=1<br />
n�<br />
i=1<br />
aisnoise,i(t)<br />
<strong>in</strong>to its underly<strong>in</strong>g n source signals s with ai(t) be<strong>in</strong>g its contribution to the<br />
sensor signal, hence its mix<strong>in</strong>g coefficient. A and s are unique except for permutation<br />
and scal<strong>in</strong>g. The functional segregation of the bra<strong>in</strong> [3] closely matches<br />
the requirement of spatially <strong>in</strong>dependent sources as assumed <strong>in</strong> spatial ICA.<br />
The term snoise(t) is the time dependent noise. Unfortunately, <strong>in</strong> fMRI the noise<br />
level is of the same order of magnitude as the signal, so it has to be taken <strong>in</strong>to<br />
account. As the noise term will depend on time, it can be <strong>in</strong>cluded as additional<br />
components <strong>in</strong>to the problem. This problem is called “under-determ<strong>in</strong>ed”<br />
or “over-complete” as the number of <strong>in</strong>dependent sources will always exceed the<br />
number of measured sensor signals x(t).<br />
Various algorithms utiliz<strong>in</strong>g higher order statistics have been proposed to solve<br />
the BSS problem. In fMRI analysis, mostly the extended Infomax (based on<br />
entropy maximisation [4, 5]) and FastICA (based on negentropy us<strong>in</strong>g fix-po<strong>in</strong>t