On gradient descent training under data augmentation with on-line noisy copies