Invariant Representation Learning