On Learning Invariant Representation for Domain Adaptation