Thank you! It seems that Hinton’s capsules learn euclidean discrepancies, too. The missing piece is that a ‘face capsule’ should also ask: “assuming there really is a face, how are its parts misaligned?”