Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity