Normalizing the training data