We propose a novel algorithms, called NGN-M with a strong theoretical convergence analysis and extensive numerical evaluations showing the robustness of our algorithm to the choice of the learning rate hyperparameter.