Preprint reports lower per-step time in neural-network tests
An arXiv preprint reports less time per iteration than classical and Krylov-subspace Levenberg-Marquardt methods in three synthetic network comparisons, while final training errors were similar.