On the Complexity of Learning Neural Networks