<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body text="#000000" bgcolor="#ffffff">
Hi,<br>
<br>
Sorry for getting back to this thread after so much time<br>
but I'm incidentally on CSP again. It could be of interest<br>
to discuss more on this and double-dipping. See below.<br>
<br>
On 11/29/2010 09:09 AM, Jakob Scherer wrote:<span
style="white-space: pre;"></span><br>
...<br>
<span style="white-space: pre;">> 1. running PCA, CSP, etc on the
whole dataset _prior_ feeding it to<br>
> a classifer: looks to me as a case of 'double-dipping', as all<br>
> trials (training and test) are used to identifiy the<br>
> components. thus all trials in the dataset given to the classifier<br>
> are actually inter-dependent. am i right there?</span><br>
...<br>
<span style="white-space: pre;">> On Thu, Nov 25, 2010 at 6:22 PM,
Emanuele Olivetti<br>
> <a class="moz-txt-link-rfc2396E" href="mailto:emanuele@relativita.com"><emanuele@relativita.com></a> wrote:<br>
>> About CSP... it is using class-labels. Isn't it?<br>
><br>
> actually no, it does not use class labels. i had a look at the<br>
> algorithm published in "2007 Liao Yao Wu Li Ieee Trans Biomed Engin<br>
> Combining Spatial Filters for the Classification of Single-Trial
EEG<br>
> in a Finger Movement Task", (where they introduce the DSP as well).<br>
></span><br>
<br>
Actually yes. I re-read Blankertz et al. 2008 (IEEE Sig.Proc.) [0]:<br>
"""In a nutshell, CSP filters maximize the variance of the spatially<br>
filtered signal under one condition while minimizing it for the<br>
other condition""" (see from p.45 on). So you need to group trials<br>
according to their class labels and do the simultaneous diagonalization<br>
of the two covariance matrices.<br>
<br>
I had a quick look to your reference and found equivalent statements<br>
and equations both for CSP and DSP.<br>
<br>
Differently from PCA, CSP and DSP might be subject to double-dipping.<br>
And they are in the case of your question.<br>
<br>
Best,<br>
<br>
Emanuele<br>
<br>
P.S.: I know there are many variants of the basic CSP algorithm but<br>
I haven't found "unsupervised" ones in the tiny fraction I met. Please<br>
keep me posted if you have different experience.<br>
<br>
[0]: <a
href="http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4408441">http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4408441</a><br>
<br>
</body>
</html>