mlsanityData to 7 Oct 2026

Priors learned from legacy reconstructions inherit undetectable overconfidence

arXiv 2607.21721 ↗First posted 23 Jul 2026stat.ML cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Where truths are scarce (e.g., seismic and medical imaging), learned priors in ill-posed inverse problems are trained on archives of legacy reconstructions---i.e., an older method's outputs---and their reported uncertainty is taken as data-driven. We show that this prior is, in the population limit, exactly the regularizer that produced its archive of posterior samples, advanced one expectation--maximization step toward the truth. While the step improves the regularizer on the directions the measurements resolve, it leaves the regularizer's assumption on the operator's blind subspace unchanged. An archive of single-best reconstructions collapses the blind interval to zero width. Neither error is detectable in practice, as truths differing only on the blind subspace share the data law, and simulation-based calibration is neutral by construction. We identify from the operator alone which directions the measurements do not inform, and, given a handful of ground-truth models, build intervals there that contain the truth as often as they claim to. We validate these findings on a two-dimensional example with closed-form predictions and in controlled experiments on seismic-imaging and groundwater-flow operators, against priors trained on the truth.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 42 of its 104 references that are in the map. First posted on arXiv 23 Jul 2026.

Builds on

42 of 104 references.

References · 42

  1. 2020Denoising Diffusion Probabilistic Modelscited by 20,851
  2. 2020Denoising Diffusion Implicit Modelscited by 9,655
  3. 2022Diffusion Posterior Sampling for General Noisy Inverse Problemscited by 1,375
  4. 2018On the Spectral Bias of Neural Networkscited by 1,321
  5. 2018fastMRI: An Open Dataset and Benchmarks for Accelerated MRIcited by 464
  6. 2021Robust Compressed Sensing MRI with Deep Generative Priorscited by 284
  7. 2023Self-Consuming Generative Models Go MADcited by 239
  8. 2019On instabilities of deep learning in image reconstruction - Does AI come at a cost?cited by 211
  9. 2017Solving ill-posed inverse problems using iterative deep neural networkscited by 145
  10. 2018Validating Bayesian Inference Algorithms with Simulation-Based Calibrationcited by 118
  11. 2023Ambient Diffusion: Learning Clean Distributions from Corrupted Datacited by 108
  12. 2022Image-to-Image Regression with Distribution-Free Uncertainty Quantification and Applications in Imagingcited by 107
30 more
  1. 2018NETT: Solving Inverse Problems with Deep Neural Networkscited by 89
  2. 2018Deep Bayesian Inversioncited by 60
  3. 2024Learning Diffusion Priors from Observations by Expectation Maximizationcited by 60
  4. 2023How to Trust Your Diffusion Model: A Convex Optimization Approach to Conformal Risk Controlcited by 50
  5. 2020The troublesome kernel -- On hallucinations, no free lunches and the accuracy-stability trade-off in inverse problemscited by 40
  6. 2023Conditional score-based diffusion models for Bayesian inference in infinite dimensionscited by 37
  7. 2019HINT: Hierarchical Invertible Neural Transport for Density Estimation and Bayesian Inferencecited by 31
  8. 2023A prior regularized full waveform inversion using generative diffusion modelscited by 29
  9. 2022Sensing Theorems for Unsupervised Learning in Linear Inverse Problemscited by 28
  10. 2021A Trust Crisis In Simulation-Based Inference? Your Posterior Approximations Can Be Unfaithfulcited by 27
  11. 2014Likelihood-informed dimension reduction for nonlinear inverse problemscited by 26
  12. 2020Representational aspects of depth and conditioning in normalizing flowscited by 25
  13. 2022Reliable amortized variational inference with physics-based latent distribution correctioncited by 25
  14. 2022PatchNR: Learning from Very Few Images by Patch Normalizing Flow Regularizationcited by 22
  15. 2024Rate of Model Collapse in Recursive Trainingcited by 22
  16. 2014Optimal low-rank approximations of Bayesian linear inverse problemscited by 21
  17. 2024An Expectation-Maximization Algorithm for Training Clean Diffusion Models from Corrupted Observationscited by 20
  18. 2021Preconditioned training of normalizing flows for variational inference in inverse problemscited by 19
  19. 2016Maximum a Posteriori Estimators as a Limit of Bayes Estimatorscited by 12
  20. 2021Deep Bayesian inference for seismic imaging with taskscited by 12
  21. 2024ASPIRE: Iterative Amortized Posterior Inference for Bayesian Inverse Problemscited by 11
  22. 2021Learning by example: fast reliability-aware seismic imaging with normalizing flowscited by 10
  23. 2023WISE: full-Waveform variational Inference via Subsurface Extensionscited by 10
  24. 2024Prediction-Centric Uncertainty Quantification via MMDcited by 8
  25. 2024Do Bayesian imaging methods report trustworthy probabilities?cited by 6
  26. 2024Tackling the Problem of Distributional Shifts: Correcting Misspecified, High-Dimensional Data-Driven Priors for Inverse Problemscited by 5
  27. 2025A Recovery Theory for Diffusion Priors: Deterministic Analysis of the Implicit Prior Algorithmcited by 3
  28. 2024Legitimate ground-truth-free metrics for deep uncertainty classification scoringcited by 1
  29. 2026On the role of memorization in learned priors for geophysical inverse problemscited by 1
  30. 2026OpenSeisML: Open Large-Scale Real Seismic and well-log Dataset for Generative AIcited by 1

Built on by

Not cited yet.