The Efficiency and the Robustness of Natural Gradient Descent Learning Rule