tl;dr deep learning needs data to learn invariances, "capsules" build in invariances to 3D rotation... somehow...