← All AI/ML open problems

Bayesian learning · Statistical inference / Open question

When Can Bayesian Confidence Be Trusted?

A precise posterior can still be confidently wrong. Useful Bayesian learning needs uncertainty that remains informative when both the model and inference are approximate.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-19

Seed catalog · Showing the bundled record.

Problem statement

Use Gaussian regression with a width-mm two-layer tanh mean fθf_\theta, compact parameter support and bounded inputs. The true mean is fθ⋆+δf_{\theta_\star}+\delta with ∥δ∥∞≤η\|\delta\|_\infty\leq\eta, and an approximate posterior qq has computable variational or sampling residuals after TT steps. For the predictive mean functional F(f)=EX∼PXf(X)F(f)=\mathbb E_{X\sim P_X}f(X), construct intervals I(D)I(D) satisfying inf⁡θ⋆,δPr⁡(F(fθ⋆+δ)∈I(D))≥1−α\inf_{\theta_\star,\delta}\Pr(F(f_{\theta_\star}+\delta)\in I(D))\geq1-\alpha. Bound interval width in sample size, prior sensitivity, misspecification, calibration labels and inference work.

Assumptions and scope

  • Specify the estimand and distinguish Bayesian posterior credibility from repeated-sampling coverage; a pseudo-true parameter need not equal a physically meaningful ground truth.
  • Fix a quantitative misspecification class, prior support and observation/selection model. Small unconditional distribution distance alone need not ensure posterior stability.
  • State the variational or sampling approximation and its computational error, together with any calibration data or generalized-posterior temperature selection.

What would settle it?

A nonvacuous finite-sample predictive coverage/risk bound and calibration rule under explicit stability conditions, with necessity examples and quantified statistical-versus-computational error.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

Misspecified variational Bernstein–von Mises results, generalized Bayes calibration and posterior-brittleness constructions are established. The target is a finite-data, computation-aware reliability guarantee for the specified nonlinear predictive model, not generic robustness of all Bayesian inference. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Wang and Blei — Variational Bayes under Model Misspecification ↗partial theory · source checked 2026-11-19

    Primary abstract checked: asymptotic normality around a KL-minimizing pseudo-true parameter under assumptions; prediction and variational/misspecification errors are analyzed. This does not imply finite-sample coverage for arbitrary neural posteriors.

  2. Baek, Aquino and Mukherjee — Generalized Bayes Approach to Inverse Problems with Model Misspecification ↗partial theory · source checked 2026-11-19

    Primary abstract checked: Gibbs posteriors, loss comparison, regularization calibration and theoretical properties for inverse problems. Robust generalized updates and calibration methods already exist.

  3. Owhadi, Scovel and Sullivan — Brittleness of Bayesian Inference Under Finite Information in a Continuous World ↗partial theory · source checked 2026-11-19

    Primary abstract checked: sharp sensitivity constructions show large posterior errors can persist under small model perturbations and finite-precision observations. Distributional closeness alone is not a universal posterior-stability condition.