issue-37

In an earlier letter, I wrote about the challenge of robustness: A learning algorithm that performs well on test data often doesn’t work well in a practical production environment because the real world turns out to be different than the test set.