Intersubject consistency of discriminative patterns
To get an impression of the consistency of the discriminative patterns
across participants, theweight maps obtained for each participant
at RFE iteration 12, which yielded the highest mean accuracy over participants,
were averaged (Fig. 4 left). Themap emphasizes voxelswith a
stronger contribution in more participants. Contributing voxelswere located
throughout the brain and did not single out specific structures as
specifically important. Also, voxels in the cerebellum and subcortical
structures (colliculus superior, striatum, thalamus, hippocampus, and
amygdala) contributed to the classification in at least some participants.
As a more objective measure of spatial consistency, we performed
intersubject classification using leave-one-participant-out cross validation
with examples fromimmediate and long term recall trials, and initial
feature reduction to 5000 voxels, in combination with recursive
feature elimination. The best accuracy of 65.3% was significantly above