Representation Learning via Invariant Causal Mechanisms