Robustness to distributional shift
Concrete Problems in AI Safety — refined
The problem of ensuring ML systems perform reasonably, and know when they are performing badly, when deployed on a test distribution that differs from the training distribution.
A model trained in one setting will eventually meet a world that looks different, and its confidence may not survive the move. Robustness to distributional shift asks that systems either keep working there, or at least notice that they are failing.
Stating the problem plainly gives way to a survey of the responses it inspired -- worth lingering on for how differently they treat the same worry, some by training on more variety, one by borrowing the language of contracts and causal structure.
The problem: performing reasonably, or knowing when you aren't
Robustness to distributional shift asks that a machine learning system either perform reasonably well, or at least recognize that it is performing badly, once deployed on a test distribution $p^*$ that differs from the training distribution $p_0$ it was fit on. Concrete Problems treats this as one of its five core research problems, noting a terminological quirk of its own text: the printed section heading reads "Robustness to Distributional Change," while the abstract and body use "distributional shift" interchangeably with it (concrete-problems, §"2 Overview of Research Problems", p. 3; §"7 Robustness to Distributional Change", p. 16). The paper positions the problem as a genuine, longstanding statistical difficulty, not a novel one it invented, and spends more of its space reviewing prior work here than in the paper's other, less-studied problem areas.
Three families of response, and the children that carry them
This concept is a genuine hub of the paper's distributional-shift section, organizing five distinct child concepts along two different axes of response. The statistical-relaxation family — Covariate shift assumption, Partially specified models, and Training on multiple distributions — keeps the classical estimation framing and asks how much of the joint distribution across train and test genuinely needs to be assumed fixed, ranging from covariate shift's single strong invariant down to training-on-multiple-distributions' assumption-free empirical wager. A second family reframes the problem conceptually rather than statistically: the edge Machine learning with contracts — reframes as a specification failure → Robustness to distributional shift shows how Machine learning with contracts recasts the whole difficulty as a broken software contract — the implicit, brittle promise that train and test distributions match — while its sibling Counterfactual reasoning supplies the causal, "what would have happened had the world been different" gloss on the same underlying question. Both reframings sit in the Reframing distributional shift as broken contracts and causal structure theme, and the concept as a whole anchors that theme plus Weakening statistical assumptions to survive distribution shift, sitting in the Before the fact or after: when a safety property gets enforced connective theme and giving its name to the entire Distributional shift: relax the statistics, or rewrite the contract supertheme. What ties all five children together is a shared refusal to accept the field's default, unexamined assumption — that deployment will resemble training — at face value.