Generalization Guarantees for Neural Networks via Harnessing the Low-rank Structure of the Jacobian