pith. sign in

arxiv: 2605.18655 · v1 · pith:VWF7WXGJnew · submitted 2026-05-18 · 📊 stat.ME · astro-ph.IM

Self-Supervised Conformal Prediction with Equivariant Bootstrapping for Image Uncertainty Quantification

Henry J. Aldridge , Tob\'ias I. Liaudat , Marcelo Pereyra , Jason D. McEwen This is my paper

Pith reviewed 2026-05-20 08:19 UTC · model grok-4.3

classification 📊 stat.ME astro-ph.IM
keywords uncertainty quantificationconformal predictionequivariant bootstrappingself-supervised learningweak lensinginverse problemsmass mappingimage reconstruction
0
0 comments X

The pith

A self-supervised approach using equivariant bootstrapping and conformal prediction quantifies uncertainty in image reconstructions without ground truth data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a method for uncertainty quantification in inverse imaging problems where multiple solutions exist due to noise and ill-posed measurements. It first applies equivariant bootstrapping to exploit inherent data symmetries such as rotations and reflections, producing initial heuristic coverage estimates. These estimates are then refined through a conformal prediction calibration performed entirely in a self-supervised way. This design eliminates reliance on ground truth data for calibration, which is often unavailable and can introduce biases from distribution shifts between simulated and real observations.

Core claim

By generating heuristic coverages via equivariant bootstrapping on data symmetries and refining them with self-supervised conformal prediction, the method achieves accurate uncertainty quantification for reconstructed images in applications like weak lensing mass-mapping without any need for ground truth calibration data.

What carries the argument

Equivariant bootstrapping that exploits symmetries to create heuristic coverages, refined by a self-supervised conformal prediction calibration step.

If this is right

  • Uncertainty estimates become available for scientific imaging tasks where ground truth data cannot be obtained or simulated.
  • Downstream analyses such as cosmological inference avoid biases introduced by model-dependent calibration simulations.
  • The approach extends to any inverse problem whose data exhibit exploitable geometric symmetries.
  • Calibration remains valid even across shifts between the observed data distribution and any auxiliary data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The technique could be tested on medical or astronomical images with weaker symmetries to measure how much symmetry strength is required for reliable calibration.
  • Combining the method with existing reconstruction networks might allow joint optimization of the image estimate and its uncertainty map.
  • If heuristic coverages prove stable, the self-supervised step could reduce the overall computational budget compared with simulation-based calibration.

Load-bearing premise

Data symmetries under rotations, reflections and similar transformations remain strong enough and are preserved by the measurement operator to produce useful heuristic coverages.

What would settle it

Applying the method to data or a measurement operator where symmetries have been deliberately broken and checking whether the resulting coverage probabilities deviate significantly from the target levels.

Figures

Figures reproduced from arXiv: 2605.18655 by Henry J. Aldridge, Jason D. McEwen, Marcelo Pereyra, Tob\'ias I. Liaudat.

Figure 1
Figure 1. Figure 1: The weak lensing mass-mapping inverse imaging problem. Ground truth corresponds to simulated convergence maps. Noisy shear map observations are visualised in terms of real and imaginary components. The convergence map is reconstructed from the observed shear with Kaiser-Squires. the prediction sets constructed from (6) are approximately equivalent to those constructed from the image-space MSE. In the gener… view at source ↗
Figure 2
Figure 2. Figure 2: Coverage test performed with and without self-supervised conformalisation with SURE. Unconfor￾malised corresponds to performing only equivariant or parametric bootstrapping. its inability to capture adequate reconstruction variability. 4. CONCLUSION We present a general method for performing calibrated UQ in inverse imaging problems that is entirely self￾supervised, alleviating the need to rely on ground t… view at source ↗
read the original abstract

Inverse problems are ubiquitous in modern scientific studies and involve recovering an underlying signal from noisy observations often transformed by a measurement operator. These problems are frequently ill-posed, particularly in imaging, leading to multiple plausible solutions and considerable uncertainty in reconstructed images. In fields like the physical and biological sciences, accurate uncertainty quantification (UQ) is critical for trustworthy scientific analyses and confident diagnoses. Current UQ methods for imaging often fall short; they can be inaccurate, or require unavailable or difficult-to-acquire ground truth data for calibration, which can introduce hidden biases due to distribution shifts between calibration and observed data. We introduce a UQ approach that leverages equivariant bootstrapping to generate heuristic coverages by exploiting data symmetries. We then refine these coverages through a conformal prediction calibration step, while crucially employing a self-supervised approach to avoid the need for ground truth calibration data. We demonstrate this method with weak lensing mass-mapping, where we aim to reconstruct the convergence field from shear measurements of distant galaxies weakly-lensed by gravitational fields. Mass-mapping in particular benefits from the self-supervised approach, as simulating calibration data is expensive and relies on specific cosmological models that could introduce biases in downstream cosmological inference tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces a UQ method for inverse imaging problems that generates heuristic coverages via equivariant bootstrapping exploiting data symmetries (rotations, reflections), then refines them with a self-supervised conformal prediction calibration step that avoids ground-truth data. The approach is demonstrated on weak-lensing mass mapping, reconstructing the convergence field from shear measurements while aiming to prevent biases from simulated calibration sets that depend on specific cosmological models.

Significance. If the central construction is valid, the method would offer a practical route to distribution-free UQ in scientific imaging domains where ground truth is unavailable or expensive to simulate. By combining symmetry-based bootstrapping with self-supervised conformal calibration, it directly targets the distribution-shift problem highlighted in the abstract and could improve reliability of downstream cosmological inference. The self-supervised framing and explicit use of equivariance are strengths that distinguish it from standard conformal or bootstrap UQ pipelines.

major comments (1)
  1. [Methods (equivariant bootstrapping and measurement operator)] The load-bearing step is the claim that equivariant bootstrapping produces heuristic coverages that conformal prediction can meaningfully refine without ground truth. Because the shear-to-convergence operator is a non-local Fourier multiplier (k_i k_j / |k|^2), the manuscript must show that this operator commutes with the symmetry group actions used for bootstrapping; otherwise the bootstrap replicates lose exchangeability with the test point and the conformal quantile no longer guarantees marginal coverage. This point should be addressed with a short invariance proof or explicit numerical check in the methods section describing the operator and the bootstrap procedure.
minor comments (2)
  1. Figure captions should explicitly state whether the displayed intervals are heuristic coverages before or after the self-supervised conformal step.
  2. The definition of the nonconformity score used in the self-supervised calibration should be written as an equation rather than described only in prose.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for your thoughtful review and for recognizing the potential of our self-supervised conformal prediction approach. We address the major comment point by point below.

read point-by-point responses

  1. Referee: The load-bearing step is the claim that equivariant bootstrapping produces heuristic coverages that conformal prediction can meaningfully refine without ground truth. Because the shear-to-convergence operator is a non-local Fourier multiplier (k_i k_j / |k|^2), the manuscript must show that this operator commutes with the symmetry group actions used for bootstrapping; otherwise the bootstrap replicates lose exchangeability with the test point and the conformal quantile no longer guarantees marginal coverage. This point should be addressed with a short invariance proof or explicit numerical check in the methods section describing the operator and the bootstrap procedure.

    Authors: We appreciate this observation and agree that a demonstration of the operator's commutation with the symmetry group is necessary to rigorously support the exchangeability assumption underlying the conformal calibration. In the revised version of the manuscript, we will include a brief invariance proof in the Methods section. The shear-to-convergence operator, expressed as a Fourier-space multiplier involving terms like (k_1^2 - k_2^2)/|k|^2 and 2k_1 k_2 / |k|^2, is equivariant under rotations and reflections because these transformations rotate the wavevector k accordingly, preserving the form of the multiplier. We will explicitly show that for a symmetry transformation g in the group (rotations by 90 degrees and reflections), the operator A satisfies A(g * shear) = g * (A * shear), ensuring the bootstrapped samples are exchangeable. We will also add a numerical check by verifying that the reconstructed convergence fields under bootstrapped symmetries match the transformed originals within numerical precision. This addition will strengthen the theoretical foundation without altering the core results. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent bootstrapping and conformal steps

full rationale

The paper's core construction generates heuristic coverages via equivariant bootstrapping that exploits data symmetries, then applies a separate self-supervised conformal calibration step to refine them without ground-truth data. No quoted equations or sections reduce the final coverage guarantee to a fitted parameter or self-citation by construction; the symmetry exploitation and conformal quantile are presented as distinct operations whose validity depends on external assumptions about the measurement operator rather than tautological redefinition. The abstract and high-level description contain no self-definitional loops, fitted-input predictions, or load-bearing self-citations that collapse the claimed result to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that imaging data possesses exploitable equivariant symmetries and that self-supervised conformal calibration can produce reliable coverage without external ground truth; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)
  • domain assumption Imaging data possesses equivariant symmetries (rotations, reflections, etc.) that are preserved under the measurement operator and can be used to generate heuristic coverages via bootstrapping.
    Invoked when the abstract states that equivariant bootstrapping generates heuristic coverages by exploiting data symmetries.
  • domain assumption Self-supervised conformal calibration can refine heuristic coverages to achieve valid uncertainty statements without introducing new biases from distribution shift.
    Central to the claim that the method avoids the need for ground truth calibration data.

pith-pipeline@v0.9.0 · 5757 in / 1648 out tokens · 41230 ms · 2026-05-20T08:19:32.432201+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · 2 internal anchors

  1. [1]

    Deep Learning Techniques for Inverse Problems in Imaging,

    G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, and R. Willett, “Deep Learning Techniques for Inverse Problems in Imaging,”IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 39–56, May 2020

    work page 2020

  2. [2]

    Do Bayesian imaging methods report trustworthy proba- bilities?

    D. Y . W. Thong, C. K. Mbakam, and M. Pereyra, “Do Bayesian imaging methods report trustworthy proba- bilities?”arXiv preprint arXiv:2405.08179, May 2024

    work page arXiv 2024

  3. [3]

    A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification

    A. N. Angelopoulos and S. Bates, “A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification,”arXiv preprint arXiv:2107.07511, Dec. 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    Theoretical Foundations of Conformal Prediction

    A. N. Angelopoulos, R. F. Barber, and S. Bates, “Theoretical foundations of conformal prediction,”arXiv preprint arXiv:2411.11824, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Self-supervised Conformal Prediction for Uncertainty Quan- tification in Imaging Problems,

    J. M. Everink, B. T. Amougou, and M. Pereyra, “Self-supervised Conformal Prediction for Uncertainty Quan- tification in Imaging Problems,” inScale Space and V ariational Methods in Computer Vision. Springer Nature Switzerland, 2025, pp. 108–118

    work page 2025

  6. [6]

    Self-supervised conformal prediction for uncertainty quan- tification in Poisson imaging problems,

    B. Tamo Amougou, M. Pereyra, and B. Pascal, “Self-supervised conformal prediction for uncertainty quan- tification in Poisson imaging problems,” inProceedings of IEEE Statistical Signal Processing Workshop, Jun. 2025

    work page 2025

  7. [7]

    Estimation of the Mean of a Multivariate Normal Distribution,

    C. M. Stein, “Estimation of the Mean of a Multivariate Normal Distribution,”The Annals of Statistics, vol. 9, no. 6, pp. 1135–1151, Nov. 1981

    work page 1981

  8. [8]

    Image-to-Image Regression with Distribution-Free Uncertainty Quantification and Applications in Imaging,

    A. N. Angelopoulos, A. P. Kohli, S. Bates, M. Jordan, J. Malik, T. Alshaabi, S. Upadhyayula, and Y . Ro- mano, “Image-to-Image Regression with Distribution-Free Uncertainty Quantification and Applications in Imaging,” inProceedings of the 39th International Conference on Machine Learning. PMLR, Jun. 2022, pp. 717–730

    work page 2022

  9. [9]

    Uncertainty quantification for fast reconstruction methods using augmented equivariant bootstrap: Application to radio interferometry,

    M. Cherif, T. I. Liaudat, J. Kern, C. Kervazo, and J. Bobin, “Uncertainty quantification for fast reconstruction methods using augmented equivariant bootstrap: Application to radio interferometry,” inNEURIPS 2024 ML4PS : Machine Learning and the Physical Sciences Workshop, Dec. 2024

    work page 2024

  10. [10]

    Distribution-free uncertainty quantification for inverse problems: Application to weak lensing mass mapping,

    H. Leterme, J. Fadili, and J.-L. Starck, “Distribution-free uncertainty quantification for inverse problems: Application to weak lensing mass mapping,”Astronomy & Astrophysics, vol. 694, p. A267, Feb. 2025. 2As implemented in the DeepInverse library:https://deepinv.github.io/deepinv/api/stubs/deepinv.physics. adjoint_function.html 44th International Works...

    work page 2025

  11. [11]

    A plug-and-play approach with fast uncertainty quan- tification for weak lensing mass mapping,

    H. Leterme, A. Tersenov, J. Fadili, and J.-L. Starck, “A plug-and-play approach with fast uncertainty quan- tification for weak lensing mass mapping,” Mar. 2026

    work page 2026

  12. [12]

    Equivariant Bootstrapping for Uncertainty Quantification in Imaging Inverse Problems: 27th International Conference on Artificial Intelligence and Statistics 2024,

    J. Tachella and M. Pereyra, “Equivariant Bootstrapping for Uncertainty Quantification in Imaging Inverse Problems: 27th International Conference on Artificial Intelligence and Statistics 2024,”Proceedings of Ma- chine Learning Research, vol. 238, pp. 4141–4149, Apr. 2024

    work page 2024

  13. [13]

    Distribution-free, Risk-controlling Prediction Sets,

    S. Bates, A. Angelopoulos, L. Lei, J. Malik, and M. Jordan, “Distribution-free, Risk-controlling Prediction Sets,”Association for Computing Machinery, vol. 68, no. 6, pp. 43:1–43:34, Sep. 2021

    work page 2021

  14. [14]

    Weak Lensing for Precision Cosmology,

    R. Mandelbaum, “Weak Lensing for Precision Cosmology,”Annual Review of Astronomy and Astrophysics, vol. 56, pp. 393–433, Sep. 2018

    work page 2018

  15. [15]

    Euclid Definition Study Report,

    R. Laureijs et al., “Euclid Definition Study Report,” Oct. 2011

    work page 2011

  16. [16]

    LSST: From Science Drivers to Reference Design and Anticipated Data Products,

    ˇZ. Ivezi ´c et al., “LSST: From Science Drivers to Reference Design and Anticipated Data Products,”The Astrophysical Journal, vol. 873, no. 2, p. 111, Mar. 2019

    work page 2019

  17. [17]

    Probabilistic mass- mapping with neural score estimation,

    B. Remy, F. Lanusse, N. Jeffrey, J. Liu, J.-L. Starck, K. Osato, and T. Schrabback, “Probabilistic mass- mapping with neural score estimation,”Astronomy & Astrophysics, vol. 672, p. A51, Apr. 2023

    work page 2023

  18. [18]

    Generative modelling for mass- mapping with fast uncertainty quantification,

    J. J. Whitney, T. I. Liaudat, M. A. Price, M. Mars, and J. D. McEwen, “Generative modelling for mass- mapping with fast uncertainty quantification,”Monthly Notices of the Royal Astronomical Society, vol. 542, no. 3, pp. 2464–2479, Sep. 2025

    work page 2025

  19. [19]

    Sparse Bayesian mass mapping with uncertainties: Peak statistics and feature locations,

    M. A. Price, J. D. McEwen, X. Cai, and T. D. Kitching (for the LSST Dark Energy Science Collaboration), “Sparse Bayesian mass mapping with uncertainties: Peak statistics and feature locations,”Monthly Notices of the Royal Astronomical Society, vol. 489, no. 3, pp. 3236–3250, Nov. 2019

    work page 2019

  20. [20]

    Wiener,Extrapolation, Interpolation, and Smoothing of Stationary Time Series: With Engineering Appli- cations

    N. Wiener,Extrapolation, Interpolation, and Smoothing of Stationary Time Series: With Engineering Appli- cations. The MIT Press, Aug. 1949

    work page 1949

  21. [21]

    Weak Gravitational Lensing,

    M. Bartelmann and P. Schneider, “Weak Gravitational Lensing,”Physics Reports, vol. 340, no. 4-5, pp. 291– 472, Jan. 2001

    work page 2001

  22. [22]

    Mapping the Dark Matter with Weak Gravitational Lensing,

    N. Kaiser and G. Squires, “Mapping the Dark Matter with Weak Gravitational Lensing,”The Astrophysical Journal, vol. 404, p. 441, Feb. 1993

    work page 1993

  23. [23]

    High resolution weak lensing mass mapping combining shear and flexion,

    F. Lanusse, J.-L. Starck, A. Leonard, and S. Pires, “High resolution weak lensing mass mapping combining shear and flexion,”Astronomy & Astrophysics, vol. 591, p. A2, Jul. 2016

    work page 2016

  24. [24]

    Sparse Bayesian mass mapping with uncertainties: Local credible intervals,

    M. A. Price, X. Cai, J. D. McEwen, M. Pereyra, T. D. Kitching, and LSST Dark Energy Science Collabo- ration, “Sparse Bayesian mass mapping with uncertainties: Local credible intervals,”Monthly Notices of the Royal Astronomical Society, vol. 492, no. 1, pp. 394–404, Feb. 2020

    work page 2020

  25. [25]

    Sparse Bayesian mass-mapping with uncertain- ties: Full sky observations on the celestial sphere,

    M. A. Price, J. D. McEwen, L. Pratley, and T. D. Kitching, “Sparse Bayesian mass-mapping with uncertain- ties: Full sky observations on the celestial sphere,”Monthly Notices of the Royal Astronomical Society, vol. 500, no. 4, pp. 5436–5452, Dec. 2020

    work page 2020

  26. [26]

    Sparse Bayesian mass mapping with uncertainties: Hypothesis testing of structure,

    M. A. Price, J. D. McEwen, X. Cai, T. D. Kitching, C. G. R. Wallis, and (for the LSST Dark Energy Science Collaboration), “Sparse Bayesian mass mapping with uncertainties: Hypothesis testing of structure,”Monthly Notices of the Royal Astronomical Society, vol. 506, no. 3, pp. 3678–3690, Jul. 2021

    work page 2021

  27. [27]

    Weak-lensing mass reconstruction using sparsity and a Gaussian random field,

    J.-L. Starck, K. E. Themelis, N. Jeffrey, A. Peel, and F. Lanusse, “Weak-lensing mass reconstruction using sparsity and a Gaussian random field,”Astronomy & Astrophysics, vol. 649, p. A99, May 2021

    work page 2021

  28. [28]

    Deep learning dark matter map reconstructions from DES SV weak lensing data,

    N. Jeffrey, F. Lanusse, O. Lahav, and J.-L. Starck, “Deep learning dark matter map reconstructions from DES SV weak lensing data,”Monthly Notices of the Royal Astronomical Society, vol. 492, no. 4, pp. 5023–5029, Mar. 2020

    work page 2020

  29. [29]

    Noise reduction for weak lensing mass mapping: An application of generative adversarial networks to Subaru Hyper Suprime-Cam first-year data,

    M. Shirasaki, K. Moriwaki, T. Oogi, N. Yoshida, S. Ikeda, and T. Nishimichi, “Noise reduction for weak lensing mass mapping: An application of generative adversarial networks to Subaru Hyper Suprime-Cam first-year data,”Monthly Notices of the Royal Astronomical Society, vol. 504, no. 2, pp. 1825–1839, Apr. 2021. 44th International Workshop on Bayesian Inf...

    work page 2021

  30. [30]

    Conformal uncertainty sets for robust optimization,

    C. Johnstone and B. Cox, “Conformal uncertainty sets for robust optimization,” inProceedings of the Tenth Symposium on Conformal and Probabilistic Prediction and Applications. PMLR, Sep. 2021, pp. 72–90

    work page 2021

  31. [31]

    The Eigenvalues of Mega-dimensional Matrices,

    J. Skilling, “The Eigenvalues of Mega-dimensional Matrices,” inMaximum Entropy and Bayesian Methods: Cambridge, England, 1988, J. Skilling, Ed. Springer Netherlands, 1989, pp. 455–466

    work page 1988

  32. [32]

    A Stochastic Estimator of the Trace of the Influence Matrix for Laplacian Smoothing Splines,

    M. Hutchinson, “A Stochastic Estimator of the Trace of the Influence Matrix for Laplacian Smoothing Splines,”Communications in Statistics - Simulation and Computation, vol. 18, no. 3, pp. 1059–1076, Jan. 1989

    work page 1989

  33. [33]

    Generalized SURE for Exponential Families: Applications to Regularization,

    Y . C. Eldar, “Generalized SURE for Exponential Families: Applications to Regularization,”IEEE Transac- tions on Signal Processing, vol. 57, no. 2, pp. 471–481, Feb. 2009

    work page 2009

  34. [34]

    UNSURE: Self-supervised learning with Unknown Noise level and Stein’s Unbiased Risk Estimate,

    J. Tachella, M. Davies, and L. Jacques, “UNSURE: Self-supervised learning with Unknown Noise level and Stein’s Unbiased Risk Estimate,”arXiv preprint arXiv:2409.01985, Feb. 2025

    work page arXiv 2025

  35. [35]

    Deepinverse: A python package for solving imaging inverse problems with deep learning,

    J. Tachella, M. Terris, S. Hurault, A. Wang, L. Davy, J. Scanvic, V . Sechaud, R. V o, T. Moreau, T. Davies, D. Chen, N. Laurent, B. Monroy, J. Dong, Z. Hu, M.-H. Nguyen, F. Sarron, P. Weiss, P. Escande, M. Massias, T. Modrzyk, B. Levac, T. I. Liaudat, M. Song, J. Hertrich, S. Neumayer, and G. Schramm, “Deepinverse: A python package for solving imaging in...

    work page doi:10.21105/joss.08923 2025

  36. [36]

    COSMOS: Hubble Space Telescope Observations,

    N. Scoville, R. G. Abraham, H. Aussel, J. E. Barnes, A. Benson, A. W. Blain, D. Calzetti, A. Comastri, P. Capak, C. Carilli, J. E. Carlstrom, C. M. Carollo, J. Colbert, E. Daddi, R. S. Ellis, M. Elvis, S. P. Ewald, M. Fall, A. Franceschini, M. Giavalisco, W. Green, R. E. Griffiths, L. Guzzo, G. Hasinger, C. Impey, J.-P. Kneib, J. Koda, A. Koekemoer, O. Le...

    work page 2007

  37. [37]

    κTNG: Effect of baryonic processes on weak lensing with IllustrisTNG simulations,

    K. Osato, J. Liu, and Z. Haiman, “κTNG: Effect of baryonic processes on weak lensing with IllustrisTNG simulations,”Monthly Notices of the Royal Astronomical Society, vol. 502, no. 4, pp. 5593–5602, Apr. 2021

    work page 2021