Not all datasets are images and we need architectures that adapt to other types of data, encoding both domain specific knowledge and data specific characteristics. For instance, at the SDSC, we deal with spherical data, i.e. curved images on a sphere, but without clear borders and arbitrary orientation.