Towards Learning and Explaining Indirect Causal Effects in Neural Networks