searchlight with no target model

Adi Yaniv

unread,

May 30, 2019, 3:30:26 AM5/30/19

to CoSMoMVPA

Hi again :) (i'm really enjoying using cosmo!)

I used it successfully and managed to preform nice searchlight analysis with a target model. I suspect that my models do not represent the "real life" accurately and would like to preform a searchlight analysis without using a target matrix as a model.

what i am looking for is running over the whole brain, calculating RDMs - and then to use each time only one value from the RDM for mapping.

a quick example:

my experimental design has 3 conditions: A,B,C, which result in 3*3 dissimilarity matrix.

i would like to compute RDMs for the whole brain, and then get only the value of dissimilarity between A and C (for each voxel, or neighborhood - my data is fMRI data) - for each subject and then use cosmo_stack to put all subjects together and transform it back into a vmp map.

then i would like to do the same for the dissimilarity between B and C.

what would be the best pipeline?

thanks a lot,

Adi

Nick Oosterhof

unread,

May 30, 2019, 3:37:50 AM5/30/19

to Adi Yaniv, CoSMoMVPA

Greetings,

On Thu, 30 May 2019 at 09:30, Adi Yaniv <adi....@gmail.com> wrote:

Hi again :) (i'm really enjoying using cosmo!)

great to hear that!

I think cosmo_dissimilarity_matrix_measure is just what you need, together with cosmo_searchlight.

The output will have .samples with 3 (=3*(3-1)/2) rows corresponding to the three pairs.

In the output, .sa.targets1 and .sa.targets2 indicate the pair combination for each corresponding row in .samples

Adi Yaniv

unread,

Jun 3, 2019, 4:42:14 PM6/3/19

to CoSMoMVPA

thanks a lot! it was exactly what i was looking for and it works fine!

hope to see you at OHBM next week :)

Moritz Wurm

unread,

Jun 12, 2019, 9:20:14 AM6/12/19

to cosm...@googlegroups.com

Dear Cosmo community,

Has anyone experience with extracting confusion matrices from time
generalization data?

I am running a time generalization MVPA with 80 x 80 time bins. To get
the confusion matrices I use measure_args.output = 'winner_predictions'.
However, I cannot unflatten it (as I would do it with
measure_args.output = 'accuracy' ):

K>> cosmo_unflatten(ds_res)
Error using cosmo_isfield>single_isfield (line 115)
Struct does not have field .a.fdim

Error in cosmo_isfield (line 71)
            tf(k)=single_isfield(s,name{k},raise);

Error in cosmo_unflatten (line 134)
            cosmo_isfield(ds,{'a.fdim','samples','fa'},true);

If I choose dimension 1, I also get an error:

K>> cosmo_unflatten(ds_res,1)
Error using cosmo_unflatten>unflatten_features (line 200)
Duplicate features at #1 and #2

Error in cosmo_unflatten (line 150)
    [arr, dim_labels,dim_values]=unflatten_features(samples, ...

If I simply apply cosmo_confusion_matrix(ds), I get a single matrix
only. My workaround is a loop:

for i=1:length(ds_res.a.sdim.values{1})
            for j=1:length(ds_res.a.sdim.values{2})
                    cond_idx = find(ds_res.sa.train_time==i &
ds_res.sa.test_time==j);
                    ds=cosmo_slice(ds_res,cond_idx);
                    MAT(i,j,:,:)=cosmo_confusion_matrix(ds);
             end
end

Unfortunately, this is extremely slow. So I am wondering: is there a
better/faster way to do it?
I'd be very thankful for your suggestions!

Best wishes,
Moritz

Nick Oosterhof

unread,

Jun 15, 2019, 5:55:51 AM6/15/19

to Moritz Wurm, CoSMoMVPA

Hi Moritz,

It's a bit hard to tell what's going wrong here. Using dim=1 seems the right approach.

What is the output of cosmo_disp(rs_res)?

best,

Nick

Moritz Wurm

unread,

Jun 18, 2019, 3:54:59 AM6/18/19

to Nick Oosterhof, CoSMoMVPA

Hi Nick
Thanks for your response and sorry for the late reply.

Find below the output of cosmo_disp(rs_res). However, I noticed that it works when I compute correlation matrices instead of confusion matrices, which might be more useful for my purposes anyway. So this is not an urgent issue anymore.

Thanks!
Best,
Moritz

.samples
[ NaN
    NaN
    NaN
     :
      4
      4
      4 ]@2764800x1
.sa
.targets
    [ 1
      1
      1
      :
      4
      4
      4 ]@2764800x1
.test_time
    [ 1
       1
       1
       :
      80
      80
      80 ]@2764800x1
.train_time
    [ 1
       1
       1
       :
      80
      80
      80 ]@2764800x1
.transpose_ids
    [ 1
       2
       3
       :
      78
      79
      80 ]@2764800x1
.trialinfo
    [ 3
      3
      3
      :
      6
      6
      6 ]@2764800x1
.a
.sdim
    .labels
      { 'train_time' 'test_time' }
    .values
      { [ 0.0145         [ 0.0145
          0.0345           0.0345
          0.0545           0.0545
             :                :
            1.55             1.55
            1.57             1.57
            1.59 ]@80x1      1.59 ]@80x1 }

and this is the dataset when I use correlation matrices instead of confusion matrices:

.samples
[ 0.634
     0.306
     0.629
       :
     0.107
     0.127
    -0.226 ]@230400x1
.sa
.half1
    [ 1
      2
      3
      :
      6
      7
      8 ]@230400x1
.half2
    [ 1
      1
      1
      :
      8
      8
      8 ]@230400x1
.test_time
    [ 1
       1
       1
       :
      45
      45
      45 ]@230400x1
.train_time
    [ 1
       1
       1
       :
      80
      80
      80 ]@230400x1
.a
.sdim
    .labels
      { 'half1' 'half2' 'train_time' 'test_time' }
    .values
      { [ 1        [ 1        [ 0.0145         [ 0.0145
          2          2          0.0345           0.0345
          3          3          0.0545           0.0545
          :          :             :                :
          6          6            1.55            0.854
          7          7            1.57            0.874
          8 ]@8x1    8 ]@8x1      1.59 ]@80x1     0.894 ]@45x1 }

Reply all

Reply to author

Forward