Exact and Consistent Interpretation for Piecewise Linear Neural Networks: A Closed Form Solution