Searchlight for representational similarity analysis

Using cosmo_searchlight, run cross-validation with nearest neighbor classifier

  1. For CoSMoMVPA's copyright information and license terms, #
  2. see the COPYING file distributed with CoSMoMVPA. #

Contents

Define data

config=cosmo_config();
study_path=fullfile(config.tutorial_data_path,'ak6');

data_path=fullfile(study_path,'s01');
data_fn=fullfile(data_path,'glm_T_stats_perrun.nii');
mask_fn=fullfile(data_path,'brain_mask.nii');
targets=repmat(1:6,1,10)';
ds = cosmo_fmri_dataset(data_fn, ...
                        'mask',mask_fn,...
                        'targets',targets);

% compute average for each unique target, so that the dataset has 6
% samples - one for each target
ds=cosmo_fx(ds, @(x)mean(x,1), 'targets', 1);

models_path=fullfile(study_path,'models');
load(fullfile(models_path,'behav_sim.mat'));
load(fullfile(models_path,'v1_model.mat'));

Set measure

Set the 'measure' and 'measure_args' to use the @cosmo_target_dsm_corr_measure measure and set its parameters to so that the target_dsm is based on behav_sim.mat

% >@@>
measure = @cosmo_target_dsm_corr_measure;
measure_args = struct();
measure_args.target_dsm = behav;
% <@@<

% Enable centering the data
measure_args.center_data=true;

Run searchlight

use spherical neighborhood of 100 voxels

voxel_count=100;
% define a neighborhood using cosmo_spherical_neighborhood
% >@@>
nbrhood=cosmo_spherical_neighborhood(ds,'count',voxel_count);
% <@@<

% Run the searchlight
% >@@>
results = cosmo_searchlight(ds,nbrhood,measure,measure_args);
% <@@<

% Save the results to disc using the following command:
output_path=config.output_data_path;
cosmo_map2fmri(results, ...
            fullfile(output_path,'rsm_searchlight_behav.nii'));

+00:00:05 [####################] -00:00:00  mean size 99.8
+00:00:29 [####################] -00:00:00

Make a histogram of correlations

hist(results.samples,47)

Show some slices

cosmo_plot_slices(results);

Advanced exercise: regresion-based RSA

% Using @cosmo_target_dsm_corr_measure, investigate the relative
% contributions of the v1-model and behavioural similarity matrix.
%
% Thus, set the 'measure' and 'measure_args' to use the
% @cosmo_target_dsm_corr_measure measure and set its parameters
% so that the 'glm_dsm' option uses the 'behav' and 'v1_model' targets
% >@@>
measure = @cosmo_target_dsm_corr_measure;
measure_args = struct();
measure_args.glm_dsm = {behav, v1_model};

% <@@<
% Enable centering the data
measure_args.center_data=true;

Run searchlight

use spherical neighborhood of 100 voxels

voxel_count=100;
% define a neighborhood using cosmo_spherical_neighborhood
% >@@>
nbrhood=cosmo_spherical_neighborhood(ds,'count',voxel_count);
% <@@<

% Run the searchlight
% >@@>
glm_dsm_results = cosmo_searchlight(ds,nbrhood,measure,measure_args);
% <@@<

% Save the results to disc using the following command:
output_path=config.output_data_path;
cosmo_map2fmri(glm_dsm_results, ...
            fullfile(output_path,'rsm_searchlight_glm_behav-v1.nii'));

+00:00:06 [####################] -00:00:00  mean size 99.8
+00:01:05 [####################] -00:00:00

Show behavioural seachrlight map

figure();
cosmo_plot_slices(cosmo_slice(glm_dsm_results,1));

Show V1 searchlight map

figure();
cosmo_plot_slices(cosmo_slice(glm_dsm_results,2));


Published with MATLAB® R2016b