Fig. 4

Description of the co-smoothing procedure to compute the jointly activated brain areas using FastSRM. The algorithm runs two nested CV loops. One first outer loop executes a random split of the cohort of twelve participants into a train set and a test set, respectively composed of eight and four subjects. One inner loop executes a random split of the group of runs into a train set and a test set: respectively twelve and nine for Clips and, interchangeably seven or six for Raiders. For every turn of the nested CV loops: (a) the k = 20 spatial components specific to each subject on the train runs are computed through alternate minimization together with their shared response, which is then used to compute the individual components of test subjects on the same runs; then, (b) assuming that the same features of the train runs will be found on the test runs, we fit the individual responses of the train subjects on the test runs in order to compute their shared response; (c) test runs are then predicted through their shared response computed in (b); and, (d) the vertex-wise correlation between the predicted runs and the corresponding original data is computed. For every subject, we estimated the vertex-wise median of the correlations across runs. The vertex-wise median of the correlations across all subjects was then estimated from their individual median correlations. This final coefficient represents the similarity of activated regions across subjects for each task.