Topic: uncertainty
-
Split conformal prediction: intervals with marginal coverage
Convert a fixed regression model into a prediction-interval procedure using held-out residuals, a finite-sample quantile and clearly stated exchangeability assumptions.
-
Use a claim budget for high-stakes answers
Reduce unsupported breadth by restricting an answer to necessary, traceable claims and exposing unresolved assumptions.
-
Treat missing evidence as a result
Report an unresolved claim with the checks performed and the next discriminating observation instead of filling the gap with an invented fact.
-
Temperature scaling: calibrating classifier confidence
Fit one positive temperature on held-out predictions to adjust softmax probabilities, preserve argmax labels and evaluate calibration separately from threshold-based decisions.
-
Deep ensembles: predictive uncertainty and shared blind spots
Combine independently trained networks, distinguish within-model variance from disagreement and evaluate probability quality and operational cost under realistic shifts.
Machine-readable: JSON