“A core limitation of geometric deep learning is its assumption that all transformations are invertible, which fails to model many computational processes that destroy information.”