Text this: Approximation approaches for training neural network problems with dynamic mini-batch sub-sampled losses