Figure 3. Results on functional brain connectivity.

(A) Our model consistently provides a better fit to the data than greedy clustering, explaining the same amount of variance with 30 fewer clusters (different points were generated from different values of the hyperparameter ${\sigma }_0^2$). (B) When using our group-learned clustering to explain variance in 20 individual subjects, we consistently generalize better than the greedy clusters for cluster sizes less than 200 (^∗ p < 0.05, ^∗∗ p < 0.01). (C) A sample 172-cluster parcellation from our method. (D) Comparison between our parcels and retinotopic maps, showing a transition from eccentricity-based divisions to field map divisions.