Hardness of Learning Neural Networks under the Manifold Hypothesis Jason Wang Melanie Weber