The most basic assumption used in statistical learning theory is that
training data and test data are drawn from the same underlying distribution.
Unfortunately, in many applications, the "in-domain" test data is drawn from a
distribution that is related, but not identical, to the "out