Comparison of SRM to other multisubject approaches. Three fMRI datasets were collected as participants viewed and/or listened to the same stimulus. The data were then anatomically aligned to either Talairach (TAL) or Montreal Neurological Institute (MNI) space, giving a common coordinate system for all participants. In these examples, analyses were restricted to the time series of BOLD activity across voxels within a specific ROI. The strength of shared cognitive states during the stimulus was evaluated by attempting to identify a short movie segment in a held-out participant’s test data, based on the test data of the other participants. This was done before (TAL, MNI) or after applying an across-participant factor model: principal component analysis (PCA), independent component analysis (ICA), hyperalignment (HA) or SRM. For the methods with dimensionality reduction (PCA, ICA, SRM), k = 50 features were used. Assuming the movie segments are independent, chance accuracy is 0.001. Error bars reflect s.e.m. Adapted with permission from ref. 65, Curran Associates, Inc.