Recognition-based reconstruction.

Separately for each subject, N-fold cross-validation was performed on recognition data where N equals the number of scanning runs pooled across all of the sessions that each subject completed (i.e., 40 runs for the three subjects that completed 5 sessions each and 32 runs for the remaining subject that completed 4 sessions). For each fold, the activation patterns from N-1 runs (i.e., 39 or 31 runs) were used as training patterns and those of the remaining run served as the testing set (i.e., the trials for which the semantic component scores were predicted). In this manner, all trials iteratively contributed to both model training and model testing.