PILLAR 3 · 9 PAPERS · NO SEARCH RUN YET
Learning to defer & safety interception
When a model should abstain or hand off to a human or safeguard, from the reject option through consistent surrogate losses to sequential and multi-expert deferral.
Search string (see Methods for the per-database form)
("learning to defer" OR "learn to defer" OR "learning-to-defer" OR "selective prediction" OR "reject option" OR "abstention" OR "deferral") AND ("sequential" OR "reinforcement" OR "trajectory" OR "multi-step" OR "expert" OR "human-in-the-loop" OR "shield*" OR "runtime monitor*" OR "safety")Included papers
| Paper | Year | Venue | Contribution | Methods | Relevance | Tags |
|---|---|---|---|---|---|---|
| Two-stage learning to defer with multiple experts Mao A.; Mohri C.; Mohri M.; Zhong Y. |
2023 | NeurIPS | Predictor–rejector template with multiple experts; the formal structure PreCog's rejector follows. | surrogate-loss consistency-proof deferral | baseline RQ1 RQ2 seed | |
| Learning-to-defer for sequential medical decision-making under uncertainty Joshi S.; Parbhoo S.; Doshi-Velez F. |
2021 | arXiv | Extends L2D to sequential settings where deferral should account for future value not per-step loss. | sequential-decision deferral POMDP | baseline RQ2 seed | |
| Consistent estimators for learning to defer to an expert Mozannar H.; Sontag D. |
2020 | ICML | First consistent surrogate loss for learning to defer; the reference L2D formulation. | surrogate-loss consistency-proof deferral | baseline RQ2 seed | |
| Calibrated learning to defer with one-vs-all classifiers Verma R.; Nalisnick E. |
2022 | ICML | One-vs-all surrogate that yields calibrated deferral probabilities. | surrogate-loss calibration deferral | baseline RQ2 seed | |
| Predict responsibly: improving fairness and accuracy by learning to defer Madras D.; Pitassi T.; Zemel R. |
2018 | NeurIPS | Frames deferral as a joint objective over the model and a downstream decision-maker. | surrogate-loss deferral | baseline RQ2 seed | |
| Safe reinforcement learning via shielding Alshiekh M.; Bloem R.; Ehlers R.; Könighofer B.; Niekum S.; Topcu U. |
2018 | AAAI | A formal shield blocks unsafe actions at runtime given a temporal-logic specification. | formal-methods temporal-logic shielding | baseline RQ5 seed | |
| Selective classification for deep neural networks Geifman Y.; El-Yaniv R. |
2017 | NeurIPS | Risk-coverage trade-off for abstaining deep classifiers with a guaranteed risk bound. | reject-option risk-bounds | RQ2 seed | |
| On optimum recognition error and reject tradeoff Chow C.K. |
1970 | IEEE Trans. Information Theory | Origin of the reject option: abstain when the cost of error exceeds the cost of rejection. | reject-option decision-theory | RQ2 seed | |
| Control barrier function based quadratic programs for safety critical systems Ames A.D.; Xu X.; Grizzle J.W.; Tabuada P. |
2017 | IEEE Trans. Automatic Control | Forward invariance of a safe set via barrier functions; the control-theoretic alternative to learned interception. | control-theory barrier-functions | baseline seed |
GAPS THIS PILLAR LEAVES OPEN
- No included paper in this pillar addresses multi-agent state
- No included paper in this pillar addresses latency / jitter aware
- No included paper in this pillar addresses edge deployment
- No included paper in this pillar addresses cross-domain evaluation