A Causal View on Robustness of Neural Networks