Appendix 1—figure 5. Distribution of regularization parameters α and number of principal components (PCs) for prediction models.

For each kind of functional magnetic resonance imaging (fMRI) data (A: task fMRI; B: resting fMRI) and each kind of connectivity profile (hyperaligned fine-grained, multimodal surface matching-aligned fine-grained, coarse-grained), we summarize the distribution of model parameters (the regularization parameter, α, and the number of PCs) across all cross-validation folds (411 families × 360 regions = 147,960). The maximum number of PCs was usually 360 for coarse-grained connectivity profile and close to 876 for fine-grained connectivity profile (depending on training sample size). Models trained on fine-grained connectivity profiles (left and middle columns) tend to use less regularization (smaller αs) and more PCs compared with models trained on coarse-grained connectivity profiles (right column), especially models trained on hyperaligned fine-grained connectivity profiles. More PCs used by the model suggest that there are more dimensions in the connectivity profiles that are related to general intelligence, and less regularization suggests these PCs contain more signal relative to noise.