Readings: Trustworthy ML
5 entriesReadings on confidence, calibration, abstention, and the uncertainty apparatus. Feeds the review chapter's third section.
- Reading: Confidence, Calibration, and the Limits of the Inevitability Claim Published on: The trustworthy-ML field the mechanism argument lives in, and an honest test of its load-bearing claim. Calibration is fixable, which strengthens the confidence-as-wrong-category point; and the mainstream view that hallucination is mitigable, not inevitable, is accommodated by decoupling the thesis from hallucination rate.
- Reading: Abstention, Deferral, and the Receiver Published on: The constructive thread in trustworthy ML. Selective prediction lets a model decline to answer; learning to defer makes that choice depend on the downstream expert's competence, not the model's own uncertainty. The ML field reached action-relative admissibility independently, but stopped at the model boundary, which is where this work begins.
- Reading: What Confidence Cannot Be Published on: Testing the wrong-category argument against ML's best uncertainty tools. Confidence decomposes into aleatoric and epistemic uncertainty (neither is basis); even distribution-free conformal prediction gives only a marginal guarantee, not a per-claim warrant. The whole apparatus answers a different question than admissibility.
- Reading: The Uncertainty Apparatus, End to End Published on: Machine learning's full uncertainty machinery - MC-dropout, deep ensembles, evidential deep learning, out-of-distribution detection - each answers a distributional or novelty question, not whether a claim's basis suffices to act. A model asserts a class with 91% confidence about pure Gaussian noise. And the field's own comprehensive survey concludes that estimating predictive uncertainty is not sufficient for safe decision-making.
- Reading: The Calibration Reassurance, and When It Breaks Published on: Calibration is offered as the rescue of confidence, but it does not hold. The metric that certifies it (ECE) has numerous flaws; the headline result that modern nets are miscalibrated does not generalise to recent architectures (conceded, and it costs the thesis nothing); and calibration dissolves under dataset shift - models are confidently wrong about entirely out-of-distribution data. A fair-weather guarantee, and irreversible actions are not taken only in fair weather.