Identifying the Group-Theoretic Structure of Machine-Learned Symmetries