Aligning representations by enhancing dominant components of their kernel matrices to capture nonlinear dependencies.