Orthogonal Statistical Learning with Self-Concordant Loss