Training technique using adversarial objectives to make model representations invariant across different data domains.