Towards Provably Efficient Algorithms For Learning Neural Networks