Towards Evaluating and Training Verifiably Robust Neural Networks