* equal contribution
a short version of this paper appeared at the CDNNRIA workshop, NeurIPS 2018