A, CHA training procedure. For each parcel in the Glasser parcellation (here, region 329 with 430 cortical vertices), connectivity matrices are computed as the strength of connectivity (Pearson's correlation) between the activity timeseries of vertices within the parcel (seeds; u) and the average activation across vertices in all other parcels (targets; v). The connectivity matrices for 200 held-out, unrelated subjects are used to train the hyperalignment common model space for a given parcel. The model space is of dimensions D = [d₁…d_n] which are weighted combinations of U = [u₁…u_n], the dimensions of each subject's connectivity matrices in their native anatomical space. D represents the space where connectivity targets are best aligned across training subjects. Then, connectomes for the 915 test subjects are projected into D via subject-specific transformation matrices R, which detail a mapping from given subject i's anatomical space U_i into D. This procedure is repeated to derive R for each test subject and parcel. Then, the regional activation pattern for each parcel in anatomical space is mapped via R into the region's common space to compute the hyperaligned timeseries for each subject, which is then used to recompute connectivity matrices between hyperaligned seeds v and model dimensions d. B, Connectome granularity. Fine-scale connectomes are multivariate patterns of the connectivity weights between the timeseries responses of all individual vertices in a given parcel and the average timeseries response of each other parcel in the brain. Coarse-scale connectomes are vectors (univariate), representing the connectivity weight between the average timeseries response of a given parcel, averaged across all of its vertices, and those of each other parcel in the brain.