Page 1 of 1

about CSP filter dimension

Posted: Wed Apr 25, 2018 6:20 am
by jeonghun
I am currently working on BCI program operated by motor imagery.
While I am designing CSP filter, I found there are filter dimension.
The default value is two, What does the two filters represent? filter subsets with largest and smallest eigen value? or first and second?
Then, when I choose it four or five? What would that mean?
How should I select the number of dimension of the CSP filter?

And, is there any reference paper for that?

Re: about CSP filter dimension

Posted: Thu May 10, 2018 8:59 am
by jtlindgren
Hi,

I recommend using regularized csp trainer and reading its box documentation. The code basically does one eigendecomposition of certain matrix (see doc refs and/or code) per class, and picks k eigenvectors with largest eigenvalues per class and packs these into csp filter matrix. The appropriate amount of eigenvectors usually has to be found empirically (i.e. evaluate with users), though I havent personally encountered a situation with the usual csp+bandpower motor imagery setting where more than a few would have been useful.

Cheers,
Jussi