This is a different issue. We already make some assumptions about the relationship between training data and future data, even when not worried about interpreta

This is a different issue. We already make some assumptions about the relationship between training data and future data, even when not worried about interpretability, and these assumptions are enough for this paper's argument too.