Hardness of Learning Neural Networks under the Manifold Hypothesis