Thanks for the response, yes I tried to do a cross-subjects classification and got some results but looked noisy (0.58 classification accuracy only on some voxels following a whole brain search light)....however in the univariate FEAT GLM I get robust activations which overlap spatially across the different conditions.
I think the poor cross-subject classification may be noisy perhaps due to noise in the registration to standard space.....I thought that I could address this by trying Hyperalignment or by trying within-subject classification at the first level, but for the later it woudl be good to increase the number of images...... TO be clear each of individual had 8 blocks for each of the 2 classification targets, but each of the 4 trials within the block represents are different example withing the higher order classification target, hence this is why I though to try and estimate an independent COPE for each despite the fixed 2 s ISI...that would give me 32 COPES per classification target per subject which may be sufficient for within-subject crossvalidation? cheers ds On Fri, Aug 8, 2014 at 4:16 PM, Meng Liang <[email protected]> wrote: > Hi David, > > In my opinion, the issue of 'independence' itself is not really a problem > because they are the samples of the same classification target. However, I > do not think that pooling the PEs for original EVs and the PEs for the time > derivatives and treating them as the samples for the same classification > target is a good idea. Although they were associated with the same > event/target, they represents very different things and they are orthogonal > in mathematical sense. It's very likely that patterns in these two types of > PEs for the same classification target were very different, and if so, it > would increase the noise level by pooling them together. > > Another concern. Your way of obtaining the data samples for the > classification is a bit problematic to me. If I understand correctly, the > four trials within each mini block were the same and they were separated by > 2 sec. I assume that you defined a single EV for each trial in order to get > a PE for each trial in your GLM model? I'm not sure whether those PEs were > meaningful (or, say, carrying much useful information) given that the > trials were so close to each other and the ISI was fixed - the data > probably do not have enough power to resolve the information for each trial > - for such rapid design, a jittered ISI would have been better. > > I guess you have tried to use only the 16 COPEs (8 for each classification > target) and the results did not look good? Have you tried between-subject > classification which would give you more samples for training the > classifier (obviously whether a between-subject classification is suitable > depends on what question you are studying)? > > Best, > Meng > > ------------------------------ > Date: Fri, 8 Aug 2014 12:50:39 +0100 > > From: [email protected] > To: [email protected] > Subject: Re: [pymvpa] classification based on individual parameter > estimates from FSL > > hi, my thought is that, for instance, if 2 images (i.e. a PE and its > temporal derivative OR two basis functions) > are associated with the same fMRI event, then it appears that wont be able > to contribute independently to classification > performance becos they basically relate to the same thing. > > In my design, for each classification target I have little blocks of 4 > trials each ---with trials separated by 2 seconds. > Initially I used the averaged COPE for the mean across the 4 trial blocks, > but this gave few COPES (only 8 as there are 8 mini-blocks per > classification target per subject, > > which is little to do within subject classification. > > Hence it would be great if I could get more COPES, what am doing at the > moment is to model each trial event within each of the blocks (plus its > temporal derivative) so that I can get at least 4 COPES x 8 blocks= 32 > COPES per classification target for each subject, which I am hoping it may > be sufficient to carry out kNN or SVM within subject classification. > I am aware it is not possible to fully separate the HRF associated with > the 4 trials of each blocks (as ISI is fixed at 2 secs) > but given each of the 4 trials are of the same classification target, I > thought it should be okay. > > Of course I could try to get each PE and its temporal derivative for each > of the 4 trials of each block which would give me > 64 betas per class per subject....but I am concerned about the > independence issue outlined above > > any thoughts or suggestions welcome > > thanks! > ds > > > On Fri, Aug 8, 2014 at 11:49 AM, Meng Liang <[email protected]> > wrote: > > Hi David, > > In your case with contrasts defined as 1000, 0100, etc, the PEs and the > corresponding COPEs should be the same, so it should not make any > difference either using PEs or COPEs. But I don't really understand why you > say the PEs would not be independent. Can you explain it a bit more? > > Best, > Meng > > ------------------------------ > Date: Tue, 5 Aug 2014 16:40:39 +0100 > From: [email protected] > To: [email protected] > Subject: Re: [pymvpa] classification based on individual parameter > estimates from FSL > > > Hi Michael (and all), just a quick clarification on your previous response > to my query relating classification based on individual parameter estimates > (PEs) - you mentioned I could use the PEs associated with the temporal > derivative or even the PEs associated with a set of basis > functions....however I wonder that this PEs would not be independent (as > would be PEs obtained from different runs) > ....would it be okay to use those PEs anyways? > > A second related thing is that I have not been using the PEs exactly but > the Contrast of PEs (i.e. COPES in FSL) > associated with each EV- I have 16 EVs (8 per class) and hence obtained > COPES such that > 1000 > 0100 > 0010 > 0001 > etc > > I dont see why it would make any difference to work wit COPEs rather than > PEs, except that only with the later I could boost my dataset by using the > temporal derivatives or basis functions.... > > cheers > ds > > > > On Fri, Jul 4, 2014 at 2:33 PM, Michael Hanke <[email protected]> wrote: > > Hi, > > On Tue, Jul 01, 2014 at 12:25:40AM +0100, David Soto wrote: > > Hi Michael, indeed ..well done for germany today! :). > > Thanks for the reply and the suggestion on KNN > > I should have been more clear that for each subject I have the > > following *block > > *sequences > > ababbaabbaabbaba in TASK 1 > > ababbaabbaabbaba in TASK 2 > > > > this explains that I have 8 a-betas and 8 b-betas for each task > > AND for each subject..so if i concatenate & normalize all the beta data > > across subjects I will have 8 x 19 (subjects)= 152 beta images for class > a > > and the same for class b > > Ah, I guess you model each task with two regressors (hrf + derivative?). > You can also use a basis function set and get even more betas... > > > > then could I use SVM searchlight trained to discriminate a from b in > task1 > > betas and tested in the task2 betas? > > yes, no problem. > > Cheers, > > Michael > > PS: Off to enjoy the quarter finals ... ;-) > > > -- > Michael Hanke > http://mih.voxindeserto.de > > _______________________________________________ > Pkg-ExpPsy-PyMVPA mailing list > [email protected] > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa > > > > > -- > http://www1.imperial.ac.uk/medicine/people/d.soto/ > > _______________________________________________ Pkg-ExpPsy-PyMVPA mailing > list [email protected] > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa > > _______________________________________________ > Pkg-ExpPsy-PyMVPA mailing list > [email protected] > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa > > > > > -- > http://www1.imperial.ac.uk/medicine/people/d.soto/ > > _______________________________________________ Pkg-ExpPsy-PyMVPA mailing > list [email protected] > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa > > _______________________________________________ > Pkg-ExpPsy-PyMVPA mailing list > [email protected] > http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa > -- http://www1.imperial.ac.uk/medicine/people/d.soto/
_______________________________________________ Pkg-ExpPsy-PyMVPA mailing list [email protected] http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
