Towards Causal Representation Learning