Unsupervised Co-Learning on $G$-Manifolds Across Irreducible Representations