That's a good point. I realised a few weeks ago that I probably need
to extend the framework so that new subjects can be added more easily.
Essentially, when you compute the initial dot-products, you obtain
X1*X1'. You actually want [X1; X2]*[X1; X2]', where X2 contains data
from your new subject(s).
X = [X1; X2];
% XX = X*X';
XX = [X1*X1' X1*X2'; (X1*X2')' X2*X2'];
On 20 May 2011 09:59, Alexander Ivanov <[log in to unmask]> wrote:
> Dear SPMers!
> I have a question regarding kernel utilities... I have searched the archives
> but haven't found any info on this topic... I would just like to be sure
> that I understood its application correctly... After preprocessing steps
> (for example, using vbm8 toolbox) I have got m0wrp* images... Afterwards I
> want to generate dot-product matrices for each set of images (gm, wm, csf).
> After all I have (N)x(M) dp-matrix (calculated using "Kernel from images"
> option), where N-th row reflects N-th subject (is that correct?)... Next I'm
> training a classifier (in my case it's SVM) with training subset of
> dp-matrix and testing it using corresponding subset.
> My last question may appear to be stupid, but.... As far as I understand, if
> I want to make a class prediction of a new subject I can't use my classifier
> anymore and I have to recalculate dp-matrix again in order to get
> (N+1)x(M+1) quadratic matrix... - right?
> Please excuse my for my weird questions...
> Best Regards,