Follow up to the post "Why Don't Machine Learning Models Extrapolate?" investigating the use of a "norm" aggregator and custom featurizers