Hierarchical interpretations for neural network predictions