Beyond Tikhonov: faster learning with self-concordant losses, via iterative regularization.