Pith. sign in

REVIEW 3 major objections 5 minor 155 references

The paper argues that learned iterative reconstruction networks are all special cases of one abstract operator, so the loss, not the architecture, determines what they compute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 17:39 UTC pith:5UK5LCU4

load-bearing objection A clean operator-learning survey of unrolled networks whose practical claim about nonlinear update directions rests on an imported comparison that does not appear in this manuscript. the 3 major comments →

arxiv 2512.08444 v2 pith:5UK5LCU4 submitted 2025-12-09 eess.IV cs.LGcs.NAmath.FAmath.NAmath.OC

Learned iterative networks: An operator learning perspective

Andreas Hauptmann , Ozan \"Oktem This is my paper
classification eess.IV cs.LGcs.NAmath.FAmath.NAmath.OC MSC 65J2268T07
keywords learned iterative networksoperator learningalgorithm unrollinginverse problemslearned reconstruction operatorvariational networkslearned primal-duallearned gradient networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This chapter is a survey, but it makes a concrete scientific claim: the many learned iterative reconstruction networks in the literature are not genuinely different algorithms. If you write down the learned reconstruction operator—the recursive map obtained by truncating an iterative scheme and replacing handcrafted updates with neural networks—then learned gradient, learned proximal, variational, and learned primal-dual networks are all special cases that differ only in how the neural updating operator is parametrized. The authors split this into two independent questions: the reconstruction operator says how to compute, and the learning problem (loss plus training data) says what to compute. The practical payoff of the framework is twofold: it explains why unrolling more iterations does not mean converging to the true minimizer, and its numerical experiments indicate that for nonlinear inverse problems the choice of handcrafted update direction matters more than the network architecture.

Core claim

The core discovery is that a single abstraction—the learned reconstruction operator Rθ:Y→X defined by f_k = Λ^k_{θ_k}(g, f_0,…,f_{k−1}) for k=1,…,N—captures essentially every learned iterative network. In this abstraction the update scheme and the learning problem are decoupled. The authors show that variational networks, learned proximal networks, learned least-squares networks, and learned primal-dual networks are all obtained from this recursion by selecting a particular form for the neural updating operator: the primal-dual architecture adds a second neural operator in data space, but it reduces to a learned least-squares network when the dual update is fixed to Γ(h,h′,h″)=h′−h″. Because

What carries the argument

The machinery is the abstract unrolled recursion Rθ(g)=f_N with f_k=Λ^k_{θ_k}(g,f_0,…,f_{k−1}), in which each handcrafted update of an iterative scheme is replaced by a neural operator Λ^k_{θ_k}. The paper treats this operator formulation as independent of both discretization and learning problem. The workhorse inside the recursion is the neural updating operator, typically an image-to-image CNN; the update direction ∇Q_g(f)=A*(Af−g) (or its nonlinear analogue) is what injects the forward model. Different architectures then correspond to different placements of Λ around the update direction: variational networks use f−∇Q+Γθ(f), learned proximal networks use Γθ(f−∇Q), and learned primal-dual

Load-bearing premise

The numerical conclusions assume that matching parameter counts and training protocols isolates the effect of the update direction, and that the diffusion-approximation QPAT test stands in for nonlinear inverse problems in general.

What would settle it

Re-run the linear CT comparison using an out-of-distribution test set and per-algorithm optimized hyperparameters; if the primal-dual advantage disappears or reverses, the claim that architecture is secondary for linear problems weakens. Re-run the QPAT comparison with the radiative transfer equation instead of the diffusion approximation; if the Gauss-Newton advantage over gradient and quasi-Newton updates vanishes, the nonlinear result may be an artifact of model mismatch.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Unrolling more iterations does not by itself bring a network closer to the minimizer of the underlying variational problem; it increases capacity.
  • The loss function and training data, not the unrolled architecture, determine the statistical estimator: L2 loss approximates conditional expectation, L1 loss conditional median.
  • Learned primal-dual networks are an extension of learned gradient networks; they reduce to a learned least-squares network when the dual update is fixed.
  • For linear inverse problems, the choice among gradient-based unrolling formulations changes PSNR by less than 0.1 dB under comparable settings, whereas for nonlinear problems the update direction has a major effect.

Where Pith is reading between the lines

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

  • If the separation of 'how to compute' from 'what to compute' holds, comparisons between reconstruction networks should control for the learning problem before crediting architecture; otherwise apparent gains may be estimator changes, not design wins.
  • The nonlinear result suggests a cheap screening strategy: choose the handcrafted update direction (gradient vs Gauss-Newton vs quasi-Newton) using short greedy-trained runs before committing to expensive end-to-end training.
  • The linear result that less handcrafted structure performs slightly better in-distribution implies the benefit of unrolling for linear problems is mostly computational and generalization-oriented; a proper out-of-distribution test would reveal whether the architectural structure actually helps.
  • The framework predicts that swapping the loss function on a fixed unrolled network should move the learned estimator toward the corresponding Bayes estimator; this could be tested by training the same architecture with L2, L1, and adversarial losses on the same data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a unified operator-learning perspective for learned iterative networks in inverse problems. It defines a learned reconstruction operator and separates the 'how to compute' (architecture) from the 'what to compute' (learning problem), then surveys learned gradient networks, learned proximal networks, variational networks, learned primal-dual networks, and extensions to nonlinear problems and Newton-type methods. The main structural claim is that popular unrolled architectures—learned least squares, proximal gradient, variational networks, and learned primal-dual—are special cases of a common abstract update (Eqs. (17), (20), (42)). A short numerical study is presented to compare update formulations for linear CT and to report a comparison for nonlinear QPAT.

Significance. If the survey's framework is accepted, it provides a genuinely useful conceptual simplification: it identifies the statistical estimator as determined by the loss and training data, while the architecture determines approximation capacity and generalization behavior. The paper's strength is that the containment claims are given explicitly: Eq. (18), (19), (35), and Remark 7.1 show how specific architectures arise from the general update. The functional-analytic gradient and Hessian calculations in Appendix A are standard and correct. The survey also covers topics often omitted from such reviews, including PnP/DEQ connections, greedy training, and Newton-type unrolling. The numerical evidence is less conclusive and needs to be strengthened or qualified.

major comments (3)
  1. [§10.2 and §11] The conclusion that in nonlinear inverse problems 'the choice of update direction ... has a major effect on the performance' is supported only by the QPAT discussion in §10.2. The text states that 'the results are presented in Figure 3', but no Figure 3 appears in the manuscript; no numerical table, error bars, or experimental details are included. The forward model is also the diffusion approximation (Eq. (57)), and the entire study is imported from the preprint [108]. As written, this is not sufficient evidence for the strong conclusion in §11. The manuscript should either include the actual comparison (with variance or error bars) or clearly present it as a reported result from [108] and qualify the conclusion accordingly.
  2. [§8.2.3, Eqs. (51)–(52)] The formulas labeled BFGS are not the BFGS update formulas as defined. With H_k denoting a Hessian approximation and B_k denoting an inverse-Hessian approximation, the displayed B_k update is the DFP inverse-Hessian update, not the BFGS inverse-Hessian update; the H_k update is the BFGS inverse-Hessian update if H_k were the inverse Hessian, which contradicts the definition in Eq. (49). This is a technical error in a section that is meant to give readers correct quasi-Newton formulas and their Hilbert-space versions. Please correct the formulas and verify them against [143].
  3. [§10.1, Table 1] The linear numerical comparison reports average PSNR over 50 test samples without error bars or repeated runs. The differences between the learned gradient networks are 0.05–0.23 dB, which is likely within run-to-run variation for the same architecture. The subsequent statements that 'performance increases with decreasing structure' and that LPD 'clearly' improves visually should be supported by error bars, statistical significance, or at least qualified as preliminary. If this is intended only as an illustrative study, the wording in the text should make that explicit.
minor comments (5)
  1. [§6, Eq. (19)] Eq. (19) writes Λ_θ(f,∇Q_g(f)) := Γ_θ(f−∇Q_g(f)), but the left-hand side is a two-argument map and the right-hand side is a one-argument map. This is presumably a shorthand in which the first argument is ignored. Please state explicitly that the proximal update is obtained from the two-argument Λ_θ by Λ_θ(a,b)=Γ_θ(a−b), so that the containment of Eq. (29) in Eq. (17) is exact.
  2. [§8.2.2, Eq. (48)] There is an indexing mismatch in the special case discussed after Eq. (48): the general form uses f_k := Λ^k_{θ_k}(f_0, Δf_0, ..., f_{k-1}, Δf_{k-1}), but the no-memory special case writes f_{k+1} := Λ_{θ_k}(f_k, ...). Please align the indices for clarity.
  3. [§6.1.2, Eq. (28)] The proximal operator prox_S is stated as 'prox_S : X → R', but it is a map X → X. Please correct the type.
  4. [§6.1.1, Eq. (23)] The statement 'f_k := 0 if k < 0' is not a standard initialization condition. It likely should be 'f_k := 0 for k < 0' or should define the initial boundary values more carefully. Please clarify.
  5. [Appendix A, Eq. (62)] Equation (62) has a typographical error: the subscript on the second inner product is 'E' but should be X or Y. Also, the reference [118] in the bibliography contains the typo 'Lunz ansd Okan Oktem'; please fix.

Circularity Check

0 steps flagged

No circular derivation: the unification is explicit equation-level rewriting; the only self-cited numerical support is an imported experiment, not a fitted input.

full rationale

The paper is a survey/unification. Definition 2.1 defines a learned reconstruction operator and Eq. (8) defines unrolling; Sections 6-8 then write the known architectures (Eqs. 17, 22, 29, 35, 42) in that common form. This is an explicit equation-level comparison, not a fitted-parameter-then-predicted cycle. For example, Remark 7.1 shows learned primal-dual reduces to learned least-squares by the explicit choice Gamma(h,h',h'') = h' - h''; that is a constructional identity stated in the paper, not a hidden input to the conclusion. Section 10.1's linear comparison is an in-paper experiment; its caveats (similar-but-not-identical parameter counts, no error bars) are empirical-quality issues, not circularity. Section 10.2 imports the nonlinear QPAT comparison from the author-associated preprint [108], and the manuscript only refers to 'Figure 3' and says 'We refer to [108] for the experiments.' The concluding remark that update direction matters for nonlinear problems is therefore supported by an external (and not re-derived) study by an overlapping author. This is a self-citation burden and an evidence gap, but it is an empirical comparison rather than an analytic step that reduces by construction to its own inputs. No circular step satisfies the required standard of exhibiting Eq. X = Eq. Y or a fitted parameter renamed as a prediction; the score reflects only the minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The conceptual framework introduces no fitted constants, no new physical entities, and no ad hoc parameters. The free parameters listed belong only to the numerical illustration; the central survey claim is a reorganization of existing material.

free parameters (3)
  • step-size ω = learned, shared across iterates
    In the proximal and variational network updates (54)–(55), the step-size is learned during training; the numerical comparison depends on it, but the central survey claim does not.
  • unrolled iteration count N = 10 (5 for one LPD variant)
    Chosen by hand in Section 10.1; affects parameter count and performance, but is not part of the operator-learning framework.
  • training hyperparameters = 25,000 iterations, lr 1e-3, cosine annealing, 5-layer 32-channel ResNet
    Used for the linear numerical comparison in Section 10.1; not tied to the conceptual framework.
axioms (6)
  • domain assumption X and Y are Hilbert spaces and the forward operator A: X→Y is known and Fréchet differentiable in the nonlinear case.
    Definition 2.1 and Section 8.1: unrolled updates rely on evaluating A and A* or (∂A(f))*.
  • domain assumption Training data are IID draws from measures consistent with g = A(f) + e.
    Definition 2.1 and Section 4: the definition of a learned reconstruction operator presumes this data-generation model.
  • standard math Riesz representation theorem identifies gradients in Hilbert space.
    Used in Appendix A.1 to derive ∇Q_g(f) = (∂A(f))*(A(f)−g) and related formulas.
  • domain assumption Replacing handcrafted updates in a truncated iterative scheme by neural operators preserves the operator-learning interpretation.
    Section 5.3.1, Eq. (8): the entire survey rests on this unrolling construction.
  • domain assumption The diffusion approximation (57) adequately models QPAT for the numerical comparison.
    Section 10.2: the nonlinear numerical conclusions about update directions are obtained under this model simplification rather than the full radiative transfer model.
  • standard math Hilbert-space quasi-Newton formulas from [143] are correct.
    Section 8.2.3 relies on [143, Corollary 3.3] for the functional-analytic versions of BFGS, DFP, SR1, and PSB updates.

pith-pipeline@v1.3.0-alltime-deepseek · 41732 in / 12343 out tokens · 128567 ms · 2026-08-03T17:39:17.075392+00:00 · methodology

0 comments
read the original abstract

Learned image reconstruction has become a pillar in computational imaging and inverse problems. Among the most successful approaches are learned iterative networks, which are formulated by unrolling classical iterative optimisation algorithms for solving variational problems. While the underlying algorithm is usually formulated in the functional analytic setting, learned approaches are often viewed as purely discrete. In this survey we present a unified operator view for learned iterative networks. Specifically, we formulate a learned reconstruction operator, defining how to compute, and separately the learning problem, which defines what to compute. In this setting we present common approaches and show that many approaches are closely related in their core. We review linear as well as non-linear inverse problems in this framework and present a short numerical study to conclude.

Figures

Figures reproduced from arXiv: 2512.08444 by Andreas Hauptmann, Ozan \"Oktem.

Figure 1
Figure 1. Figure 1: Illustration of measurement setup. (Left) ground-truth, (Middle) [PITH_FULL_IMAGE:figures/full_fig_p037_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of reconstructions for the trained networks for one sample [PITH_FULL_IMAGE:figures/full_fig_p038_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of training update directions and training schemes for [PITH_FULL_IMAGE:figures/full_fig_p040_3.png] view at source ↗

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

155 extracted references · 34 canonical work pages · 5 internal anchors

  1. [1]

    Fast and flexible X-ray tomography using the ASTRA toolbox

    Wim van Aarle et al. “Fast and flexible X-ray tomography using the ASTRA toolbox”. In:Optics Express24.22 (2016), pp. 25129–25147. doi:10.1364/OE.24.02512

    work page doi:10.1364/oe.24.02512 2016

  2. [2]

    The ASTRA Toolbox: A platform for advanced algorithm development in electron tomography

    Wim van Aarle et al. “The ASTRA Toolbox: A platform for advanced algorithm development in electron tomography”. In:Ultramicroscopy157 (2015), pp. 35–47.doi:10.1016/j.ultramic.2015.05.002

    doi:10.1016/j.ultramic.2015.05.002 2015

  3. [3]

    TensorFlow: a system for large-scale machine learn- ing

    Mart ´ ın Abadi et al. “TensorFlow: a system for large-scale machine learn- ing”. In:Proceedings of the 12th USENIX conference on Operating Sys- tems Design and Implementation Pages (OSDI’16). 2016, pp. 265–283

    2016

  4. [4]

    Learned Iterative Reconstruction

    Jonas Adler. “Learned Iterative Reconstruction”. In:Handbook of Math- ematical Models and Algorithms in Computer Vision and Imaging: Math- ematical Imaging and Vision. Ed. by Ke Chen, Carola-Bibiane Sch¨ onlieb, Xue-Cheng Tai, and Laurent Younes. Springer Verlag, 2023, pp. 751–771. doi:10.1007/978-3-030-98661-2_67

    work page doi:10.1007/978-3-030-98661-2_67 2023

  5. [5]

    Learned Primal-Dual Reconstruction

    Jonas Adler and Ozan ¨Oktem. “Learned Primal-Dual Reconstruction”. In:IEEE Transactions on Medical Imaging37.6 (2018), pp. 1322–1332. doi:10.1109/TMI.2018.2799231

    arXiv 2018

  6. [6]

    Solving ill-posed inverse problems us- ing iterative deep neural networks

    Jonas Adler and Ozan ¨Oktem. “Solving ill-posed inverse problems us- ing iterative deep neural networks”. In:Inverse problems33.12 (2017), 124007 (24pp).doi:10.1088/1361-6420/aa9581

    doi:10.1088/1361-6420/aa9581 2017

  7. [7]

    Published in NeurIPS Optimal Transport workshop 2017, which was part of the 31st Conference on Neural Information Processing Systems (NeurIPS 2017)

    Jonas Adler, Axel Ringh, Ozan ¨Oktem, and Johan Karlsson.Learning to solve inverse problems using Wasserstein loss. Published in NeurIPS Optimal Transport workshop 2017, which was part of the 31st Conference on Neural Information Processing Systems (NeurIPS 2017). 2017.doi: 10.48550/arXiv.1710.10898. arXiv:1710.10898 [cs.CV]

    Pith/arXiv arXiv doi:10.48550/arxiv.1710.10898 2017

  8. [8]

    A Comprehensive Survey of Transformers in Text Recognition: Techniques, Challenges, and Future Directions

    Ali Afkari-Fahandari, Elham Shabaninia, Fatemeh Asadi-Zeydabadi, and Hossein Nezamabadi-Pour. “A Comprehensive Survey of Transformers in Text Recognition: Techniques, Challenges, and Future Directions”. In:ACM Computing Surveys(2025). Accepted for publication.doi:10. 1145/3771273

    2025

  9. [9]

    The R2D2 deep neural network series paradigm for fast precision imag- ing in radio astronomy

    Amir Aghabiglou, Chung San Chu, Arwa Dabbech, and Yves Wiaux. “The R2D2 deep neural network series paradigm for fast precision imag- ing in radio astronomy”. In:The Astrophysical Journal Supplement Se- ries273.1 (2024), p. 3.doi:10.3847/1538-4365/ad46f5. 45

    doi:10.3847/1538-4365/ad46f5 2024

  10. [10]

    Alberti, Damiana Lazzaro, Serena Morigi, Luca Ratti, and Matteo Santacesaria.Deep Unfolding Network for Nonlinear Multi-Frequency Electrical Impedance Tomography

    Giovanni S. Alberti, Damiana Lazzaro, Serena Morigi, Luca Ratti, and Matteo Santacesaria.Deep Unfolding Network for Nonlinear Multi-Frequency Electrical Impedance Tomography. 2025.doi:10.48550/arXiv.2507. 16678. arXiv:2507.16678 [math.NA]

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.16678 2025

  11. [11]

    The FEniCS Project Version 1.5

    Martin S. Alnæs et al. “The FEniCS Project Version 1.5”. In:Archive of Numerical Software3.100 (2015), pp. 9–23.doi:10.11588/ans.2015. 100.20553

    work page doi:10.11588/ans.2015 2015

  12. [12]

    Deep microlocal reconstruction for limited-angle tomography

    H´ ector Andrade-Loarca, Gitta Kutyniok, Ozan ¨Oktem, and Philipp Pe- tersen. “Deep microlocal reconstruction for limited-angle tomography”. In:Applied and Computational Harmonic Analysis59 (2022), pp. 155– 197.doi:10.1016/j.acha.2021.12.007

    work page doi:10.1016/j.acha.2021.12.007 2022

  13. [13]

    On instabilities of deep learning in image reconstruction and the potential costs of AI

    Vegard Antun, Francesco Renna, Clarice Poon, Ben Adcock, and Anders C. Hansen. “On instabilities of deep learning in image reconstruction and the potential costs of AI”. In:Proceedings of the National Academy of Sciences of the United States of America117 (2020), pp. 30088–30095. doi:10.1073/pnas.1907377117

    doi:10.1073/pnas.1907377117 2020

  14. [14]

    Optical tomography in medical imaging

    Simon Arridge. “Optical tomography in medical imaging”. In:Inverse problems15.2 (1999), R41.doi:10.1088/0266-5611/15/2/022

    doi:10.1088/0266-5611/15/2/022 1999

  15. [15]

    Networks for nonlinear dif- fusion problems in imaging

    Simon Arridge and Andreas Hauptmann. “Networks for nonlinear dif- fusion problems in imaging”. In:Journal of mathematical imaging and vision62.3 (2020), pp. 471–487.doi:10.1007/s10851-019-00901-3

    work page doi:10.1007/s10851-019-00901-3 2020

  16. [16]

    Inverse prob- lems with learned forward operators

    Simon Arridge, Andreas Hauptmann, and Yury Korolev. “Inverse prob- lems with learned forward operators”. In:Data-driven Models in Inverse Problems. Ed. by Tatiana A. Bubba. Vol. 31. Radon Series on Computa- tional and Applied Mathematics. Walter de Gruyter, 2025, pp. 73–106. doi:https://doi.org/10.1515/9783111251233-003

    work page doi:10.1515/9783111251233-003 2025

  17. [17]

    Solving inverse problems using data-driven models

    Simon Arridge, Peter Maass, Ozan ¨Oktem, and Carola-Bibiane Sch¨ onlieb. “Solving inverse problems using data-driven models”. In:Acta Numerica 28 (2019), pp. 1–174.doi:10.1017/S0962492919000059

    doi:10.1017/s0962492919000059 2019

  18. [18]

    Gradient method for concave programming, I: local results

    Kenneth J. Arrow and Leonid Hurwicz. “Gradient method for concave programming, I: local results”. In:Studies in Linear and Non-linear Programming. Stanford Mathematical Studies in the Social Sciences 11. Standford, CA: Stanford University Press, 1958

    1958

  19. [19]

    Invertible generative models for inverse problems: mitigating rep- resentation error and dataset bias

    Muhammad Asim, Max Daniels, Oscar Leong, Ali Ahmed, and Paul Hand. “Invertible generative models for inverse problems: mitigating rep- resentation error and dataset bias”. In:Proceedings of Machine Learn- ing Research: The 37th International Conference on Machine Learning (ICML 2020). Vol. 119. 2020, pp. 399–409

    2020

  20. [20]

    Adaptive Compu- tation and Machine Learning series

    Francis Bach.Learning Theory from First Principles. Adaptive Compu- tation and Machine Learning series. MIT Press, 2024

    2024

  21. [21]

    Computed tomography reconstruction using deep image prior and learned reconstruction methods

    Daniel Otero Baguer, Johannes Leuschner, and Maximilian Schmidt. “Computed tomography reconstruction using deep image prior and learned reconstruction methods”. In:Inverse Problems36.9 (2020), p. 094004. doi:10.1088/1361-6420/aba415

    work page doi:10.1088/1361-6420/aba415 2020

  22. [22]

    2014.doi: 10.48550/arXiv.1409.0473

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.Neural Ma- chine Translation by Jointly Learning to Align and Translate. 2014.doi: 10.48550/arXiv.1409.0473. arXiv:1409.0473 [cs.CL]. 46

    Pith/arXiv arXiv doi:10.48550/arxiv.1409.0473 2014

  23. [23]

    Deep equilibrium mod- els

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. “Deep equilibrium mod- els”. In:Proceedings of the 33rd International Conference on Neural In- formation Processing Systems (NIPS’19). Vol. 32. 2019, 690–701 (Article no.: 63)

    2019

  24. [24]

    The problem of the convergence of the iteratively regularized Gauss–Newton method

    Anatolii Borisovich Bakushinskii. “The problem of the convergence of the iteratively regularized Gauss–Newton method”. In:Computational Mathematics and Mathematical Physics32.9 (1992), pp. 1353–1359

    1992

  25. [25]

    A Comprehensive Survey on Transformer-Based Machine Translation: Identifying Research Gaps and Solutions for Large Language Models

    Anasua Banerjee and Debajyoty Banik. “A Comprehensive Survey on Transformer-Based Machine Translation: Identifying Research Gaps and Solutions for Large Language Models”. In:ACM Computing Surveys (2025). Accepted for publication.doi:10.1145/3773076

    work page doi:10.1145/3773076 2025

  26. [26]

    Data-driven nonsmooth optimization

    Sebastian Banert, Axel Ringh, Jonas Adler, Johan Karlsson, and Ozan ¨Oktem. “Data-driven nonsmooth optimization”. In:SIAM Journal on Optimization30.1 (2020), pp. 102–131.doi:10.1137/18M1207685

    work page doi:10.1137/18m1207685 2020

  27. [27]

    Accelerated Forward-Backward Optimization using Deep Learning

    Sebastian Banert, Jevgenjia Rudzusika, Ozan ¨Oktem, and Jonas Adler. “Accelerated Forward-Backward Optimization using Deep Learning”. In: SIAM Journal on Optimization34.2 (2024), pp. 1236–1263.doi:10 . 1137/22M1532548

    2024

  28. [28]

    An educated warm start for deep image prior- based micro CT reconstruction

    Riccardo Barbano et al. “An educated warm start for deep image prior- based micro CT reconstruction”. In:IEEE Transactions on Computa- tional Imaging8 (2022), pp. 1210–1222.doi:10 . 1109 / TCI . 2022 . 3233188

    2022

  29. [29]

    Automatic differentiation in machine learn- ing: a survey

    Atilim G¨ une¸ s Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. “Automatic differentiation in machine learn- ing: a survey”. In:Journal of Machine Learning Research18.1 (2018), pp. 5595–5637

    2018

  30. [30]

    Deep Learning Based Computed Tomography Whys and Wherefores

    Shabab Bazrafkan, Vincent Van Nieuwenhove, Joris Soons, Jan De Been- houwer, and Jan Sijbers.Deep Learning Based Computed Tomography Whys and Wherefores. 2019.doi:10.48550/arXiv.1904.03908. arXiv: 1904.03908 [eess.IV]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.03908 2019

  31. [31]

    Biomedical photoacoustic imaging

    Paul Beard. “Biomedical photoacoustic imaging”. In:Interface Focus1.4 (2011), pp. 602–631.doi:10.1098/rsfs.2011.0028

    arXiv 2011

  32. [32]

    2024.doi:10.48550/arXiv.2403.14606

    Mathieu Blondel and Vincent Roulet.The Elements of Differentiable Programming. 2024.doi:10.48550/arXiv.2403.14606. arXiv:2403. 14606 [cs.LG]

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.14606 2024

  33. [33]

    Com- pressed sensing using generative models

    Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G. Dimakis. “Com- pressed sensing using generative models”. In:Proceedings of Machine Learning Research: The 34th International Conference on Machine Learn- ing (ICML 2017). Vol. 70. 2017, pp. 537–546

    2017

  34. [34]

    Optimiz- ing Models Performance: A Comprehensive Review and Case Study of Hyperparameters Tuning

    Ikhlass Boukrouh, Faouzi Tayalati, and Abdellah Azmani. “Optimiz- ing Models Performance: A Comprehensive Review and Case Study of Hyperparameters Tuning”. In:Proceedings of Data Analytics and Man- agement (ICDAM 2024). Ed. by Abhishek Swaroop, Bal Virdee, S´ ergio Duarte Correia, and Zdzislaw Polkowski. Vol. 1302. Lecture Notes in Networks and Systems. 2...

    doi:10.1007/978- 2024

  35. [35]

    A mathematical guide to opera- tor learning

    Nicolas Boull´ e and Alex Townsend. “A mathematical guide to opera- tor learning”. In:Handbook of Numerical Analysis: Numerical Analysis Meets Machine Learning. Ed. by Siddhartha Mishra and Alex Townsend. Vol. 25. Elsevier, 2024. Chap. 3, pp. 83–125.doi:10.1016/bs.hna.2024. 05.003

    work page doi:10.1016/bs.hna.2024 2024

  36. [36]

    PINTO: Physics- informed transformer neural operator for learning generalized solutions of partial differential equations for any initial and boundary condition

    Sumanth Kumar Boya and Deepak N. Subramani. “PINTO: Physics- informed transformer neural operator for learning generalized solutions of partial differential equations for any initial and boundary condition”. In:Computer Physics Communications315 (2025), p. 109702.doi:10. 1016/j.cpc.2025.109702

    arXiv 2025

  37. [37]

    Version 0.2.5

    James Bradbury et al.JAX: composable transformations of Python+NumPy programs. Version 0.2.5. 2018

    2018

  38. [38]

    Choose a transformer: Fourier or Galerkin

    Shuhao Cao. “Choose a transformer: Fourier or Galerkin”. In:35th Con- ference on Neural Information Processing Systems (NeurIPS 2021). 2021, 24924–24940 (Article No.: 1909)

    2021

  39. [39]

    Unsupervised approaches based on optimal transport and convex anal- ysis for inverse problems in imaging

    Marcello Carioni, Subhadip Mukherjee, Hong Ye Tan, and Junqi Tang. “Unsupervised approaches based on optimal transport and convex anal- ysis for inverse problems in imaging”. In:Data-driven Models in Inverse Problems. Ed. by Tatiana A. Bubba. Vol. 31. Radon Series on Compu- tational and Applied Mathematics. De Gruyter, 2025, pp. 107–162.doi: 10.1515/97831...

    work page doi:10.1515/9783111251233-004 2025

  40. [40]

    A First-Order Primal-Dual Al- gorithm for Convex Problems with Applications to Imaging

    Antonin Chambolle and Thomas Pock. “A First-Order Primal-Dual Al- gorithm for Convex Problems with Applications to Imaging”. In:Jour- nal of Mathematical Imaging and Vision40 (2011), pp. 120–145.doi: 10.1007/s10851-010-0251-1

    doi:10.1007/s10851-010-0251-1 2011

  41. [41]

    Plug-and-play ADMM for image restoration: Fixed-point convergence and applications

    Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. “Plug-and-play ADMM for image restoration: Fixed-point convergence and applications”. In:IEEE Transactions on Computational Imaging3.1 (2016), pp. 84–98. doi:10.1109/TCI.2016.2629286

    arXiv 2016

  42. [42]

    Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems

    Tianping Chen and Hong Chen. “Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems”. In:IEEE Transactions on Neural Networks6.4 (1995), pp. 911–917.doi:10.1109/72.392253

    doi:10.1109/72.392253 1995

  43. [43]

    Insights into analysis op- erator learning: From patch-based sparse models to higher order MRFs

    Yunjin Chen, Rene Ranftl, and Thomas Pock. “Insights into analysis op- erator learning: From patch-based sparse models to higher order MRFs”. In:IEEE Transactions on Image Processing23.3 (2014), pp. 1060–1072. doi:10.1109/TIP.2014.2299065

    arXiv 2014

  44. [44]

    Variational Model- Based Deep Neural Networks for Image Reconstruction

    Yunmei Chen, Xiaojing Ye, and Qingchao Zhang. “Variational Model- Based Deep Neural Networks for Image Reconstruction”. In:Handbook of Mathematical Models and Algorithms in Computer Vision and Imag- ing: Mathematical Imaging and Vision. Ed. by Ke Chen, Carola-Bibiane Sch¨ onlieb, Xue-Cheng Tai, and Laurent Younes. Springer Verlag, 2023, pp. 879–907.doi:10...

    work page doi:10.1007/978-3-030-98661-2_57 2023

  45. [45]

    2023.doi:10

    Chun-Wun Cheng et al.Continuous U-Net: Faster, Greater and Noise- less. 2023.doi:10 . 48550 / arXiv . 2302 . 00626. arXiv:2302 . 00626 [cs.CV]. 48

    2023

  46. [46]

    To understand double descent, we need to understand VC theory

    Vladimir Cherkassky and Eng Hock Lee. “To understand double descent, we need to understand VC theory”. In:Neural Networks169 (2024), pp. 242–256.doi:10.1016/j.neunet.2023.10.014

    work page doi:10.1016/j.neunet.2023.10.014 2024

  47. [47]

    Learning Nonlinear Electrical Impedance Tomography

    Francesco Colibazzi, Damiana Lazzaro, Serena Morigi, and Andrea Samor´ e. “Learning Nonlinear Electrical Impedance Tomography”. In:Journal of Scientific Computing90 (2022), Article No.: 58.doi:10.1007/s10915- 021-01716-4

    doi:10.1007/s10915- 2022

  48. [48]

    In- termediate layer optimization for inverse problems using deep generative models

    Giannis Daras, Joseph Dean, Ajil Jalal, and Alexandros G. Dimakis. “In- termediate layer optimization for inverse problems using deep generative models”. In:Proceedings of Machine Learning Research: The 38th Inter- national Conference on Machine Learning (ICML 2021). Vol. 139. 2021, pp. 2421–2432

    2021

  49. [49]

    Restarted con- tractive operators to learn at equilibrium

    Leo Davy, Luis M. Brice˜ no-Arias, and Nelly Pustelnik. “Restarted con- tractive operators to learn at equilibrium”. In:Machine Learning Solu- tions for Inverse Problems: Part A. Ed. by Andreas Hauptmann, Bangti Jin, and Carola-Bibiane Sch¨ onlieb. Vol. 26. Handbook of Numerical Anal- ysis. Elsevier, 2025, pp. 315–340.doi:https://doi.org/10.1016/bs. hna.2...

    doi:10.1016/bs 2025

  50. [50]

    Regularization by architecture: A deep prior approach for inverse prob- lems

    S¨ oren Dittmer, Tobias Kluth, Peter Maass, and Daniel Otero Baguer. “Regularization by architecture: A deep prior approach for inverse prob- lems”. In:Journal of Mathematical Imaging and Vision62.3 (2020), pp. 456–470.doi:10.1007/s10851-019-00923-x

    work page doi:10.1007/s10851-019-00923-x 2020

  51. [51]

    Plug-and-play image reconstruc- tion is a convergent regularization method

    Andrea Ebner and Markus Haltmeier. “Plug-and-play image reconstruc- tion is a convergent regularization method”. In:IEEE Transactions on Image Processing33 (2024), pp. 1476–1486.doi:10.1109/TIP.2024. 3361218

    work page doi:10.1109/tip.2024 2024

  52. [52]

    A survey on semi-supervised learning

    Jesper E. van Engelen and Holger H. Hoos. “A survey on semi-supervised learning”. In:Machine Learning109 (2020), pp. 373–440.doi:10.1007/ s10994-019-05855-6

    2020

  53. [53]

    BCR-Net: A neural network based on the nonstandard wavelet form

    Yuwei Fan, Cindy Orozco Bohorquez, and Lexing Ying. “BCR-Net: A neural network based on the nonstandard wavelet form”. In:Journal of Computational Physics384 (2019), pp. 1–15.doi:10 . 1016 / j . jcp . 2019.02.002

    2019

  54. [54]

    A multiscale neural network based on hierarchical nested bases

    Yuwei Fan, Jordi Feliu-Fab` a, Lin Lin, Lexing Ying, and Leonardo Zepeda- N´ u˜ nez. “A multiscale neural network based on hierarchical nested bases”. In:Research in the Mathematical Sciences6.21 (2019), 28 pp.doi:10. 1007/s40687-019-0183-3

    2019

  55. [55]

    A Mul- tiscale Neural Network Based on Hierarchical Matrices

    Yuwei Fan, Lin Lin, Lexing Ying, and Leonardo Zepeda-N´ u˜ nez. “A Mul- tiscale Neural Network Based on Hierarchical Matrices”. In:Multiscale Modeling & Simulation17.4 (2019), pp. 1–15.doi:10.1137/18M1203602

    work page doi:10.1137/18m1203602 2019

  56. [56]

    Meta-learning pseudo- differential operators with deep neural networks

    Jordi Feliu-Fab` a, Yuwei Fan, and Lexing Ying. “Meta-learning pseudo- differential operators with deep neural networks”. In:Journal of Com- putational Physics408 (2020), 109309 (18 pp.)doi:10.1016/j.jcp. 2020.109309. 49

    arXiv 2020

  57. [57]

    Deep learning tomographic reconstruction through hierarchical decomposition of domain transforms

    Lin Fu and Bruno De Man. “Deep learning tomographic reconstruction through hierarchical decomposition of domain transforms”. In:Visual Computing for Industry, Biomedicine, and Art5 (2022), Article No.: 30. doi:10.1186/s42492-022-00127-y

    work page doi:10.1186/s42492-022-00127-y 2022

  58. [58]

    Jfb: Jacobian-free backpropagation for implicit networks

    Samy Wu Fung et al. “Jfb: Jacobian-free backpropagation for implicit networks”. In:Proceedings of the AAAI Conference on Artificial Intelli- gence. Vol. 36. 6. 2022, pp. 6648–6656

    2022

  59. [59]

    Deep equilibrium architectures for inverse problems in imaging

    Davis Gilton, Gregory Ongie, and Rebecca Willett. “Deep equilibrium architectures for inverse problems in imaging”. In:IEEE Transactions on Computational Imaging7 (2021), pp. 1123–1133.doi:10.1109/TCI. 2021.3118944

    arXiv 2021

  60. [60]

    Isotropic and anisotropic total variation regularization in electrical impedance to- mography

    Gerardo Gonz´ alez, Ville Kolehmainen, and Aku Sepp¨ anen. “Isotropic and anisotropic total variation regularization in electrical impedance to- mography”. In:Computers & Mathematics with Applications74.3 (2017), pp. 564–576

    2017

  61. [61]

    Learning fast approximations of sparse coding

    Karol Gregor and Yann LeCun. “Learning fast approximations of sparse coding”. In:Proceedings of the 27th International Conference on Inter- national Conference on Machine Learning (ICML’10). 2010, pp. 399– 406

    2010

  62. [62]

    Digital twins enable full-reference quality assessment of photoacoustic image reconstructions

    Janek Gr¨ ohl et al. “Digital twins enable full-reference quality assessment of photoacoustic image reconstructions”. In:The Journal of the Acous- tical Society of America158.1 (July 2025), pp. 590–601.doi:10.1121/ 10.0037188

    2025

  63. [63]

    Fourier Neural Op- erator Network for Fast Photoacoustic Wave Simulations

    Steven Guan, Ko-Tsung Hsu, and Parag V. Chitnis. “Fourier Neural Op- erator Network for Fast Photoacoustic Wave Simulations”. In:Algorithms 16.2 (2023).doi:10.3390/a16020124

    work page doi:10.3390/a16020124 2023

  64. [64]

    Variational Models and Their Combinations with Deep Learning in Medical Image Segmentation: A Survey

    Luying Gui, Jun Ma, and Xiaoping Yang. “Variational Models and Their Combinations with Deep Learning in Medical Image Segmentation: A Survey”. In:Handbook of Mathematical Models and Algorithms in Com- puter Vision and Imaging: Mathematical Imaging and Vision. Ed. by Ke Chen, Carola-Bibiane Sch¨ onlieb, Xue-Cheng Tai, and Laurent Younes. Springer Verlag, 2...

    work page doi:10.1007/978-3-030-98661- 2023

  65. [65]

    Stable architectures for deep neural networks

    Eldad Haber and Lars Ruthotto. “Stable architectures for deep neural networks”. In:Inverse problems34.1 (2017), p. 014004.doi:10.1088/ 1361-6420/aa9a90

    2017

  66. [66]

    Learning a variational network for reconstruc- tion of accelerated MRI data

    Kerstin Hammernik et al. “Learning a variational network for reconstruc- tion of accelerated MRI data”. In:Magnetic resonance in medicine79.6 (2018), pp. 3055–3071.doi:10.1002/mrm.26977

    doi:10.1002/mrm.26977 2018

  67. [67]

    GNOT: A General Neural Operator Transformer for Operator Learning

    Zhongkai Hao et al. “GNOT: A General Neural Operator Transformer for Operator Learning”. In:Journal of Machine Learning Research202 (2023). Proceedings of the 40th International Conference on Machine (ICML 2023), 12556–12569 (Article No.: 509)

    2023

  68. [68]

    Multi-scale learned iterative reconstruction

    Andreas Hauptmann, Jonas Adler, Simon Arridge, and Ozan ¨Oktem. “Multi-scale learned iterative reconstruction”. In:IEEE transactions on computational imaging6 (2020), pp. 843–856.doi:10.1109/TCI.2020. 2990299. 50

    doi:10.1109/tci.2020 2020

  69. [69]

    2025.doi:10

    Andreas Hauptmann, Leonid Kunyansky, and Jenni Poimala.Fast algo- rithms enabling optimization and deep learning for photoacoustic tomog- raphy in a circular detection geometry. 2025.doi:10 . 48550 / arXiv . 2510.24687. arXiv:2510.24687 [eess.IV]

    work page doi:10.48550/arxiv.2510.24687 2025

  70. [70]

    Convergent regularization in inverse problems and linear plug-and-play denoisers

    Andreas Hauptmann, Subhadip Mukherjee, Carola-Bibiane Sch¨ onlieb, and Ferdia Sherry. “Convergent regularization in inverse problems and linear plug-and-play denoisers”. In:Foundations of Computational Math- ematics(2024), pp. 1–34.doi:10.1007/s10208-024-09654-x

    doi:10.1007/s10208-024-09654-x 2024

  71. [71]

    2023.doi: 10.48550/arXiv.2304.01963

    Andreas Hauptmann and Jenni Poimala.Model-corrected learned primal- dual models for fast limited-view photoacoustic tomography. 2023.doi: 10.48550/arXiv.2304.01963. arXiv:2304.01963 [eess.IV]

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.01963 2023

  72. [72]

    Approximate k-space models and deep learn- ing for fast photoacoustic reconstruction

    Andreas Hauptmann et al. “Approximate k-space models and deep learn- ing for fast photoacoustic reconstruction”. In:Machine Learning for Medical Image Reconstruction (MLMIR 2018), held in Conjunction with MICCAI 2018. Ed. by Florian Knoll, Andreas Maier, and Daniel Rueck- ert. Lecture Notes in Computer Science 11074. Springer Verlag, 2018, pp. 103–111.doi...

    work page doi:10.1007/978-3-030-00129-2_12 2018

  73. [73]

    Model-Based Learning for Accelerated, Limited- View 3-D Photoacoustic Tomography

    Andreas Hauptmann et al. “Model-Based Learning for Accelerated, Limited- View 3-D Photoacoustic Tomography”. In:IEEE Transactions on Med- ical Imaging37.6 (2018), pp. 1382–1393.doi:10 . 1109 / TMI . 2018 . 2820382

    2018

  74. [74]

    Deep Resid- ual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. “Deep Resid- ual Learning for Image Recognition”. In:IEEE Conference on Computer Vision and Pattern Recognition, CVPR(2016), pp. 770–778.doi:10. 1109/CVPR.2016.90

    2016

  75. [75]

    Do- main independent post-processing with graph U-nets: applications to electrical impedance tomographic imaging

    William Herzberg, Andreas Hauptmann, and Sarah J Hamilton. “Do- main independent post-processing with graph U-nets: applications to electrical impedance tomographic imaging”. In:Physiological measure- ment44.12 (2023), p. 125008.doi:10.1088/1361-6579/ad0b3d

    work page doi:10.1088/1361-6579/ad0b3d 2023

  76. [76]

    Graph convolutional networks for model-based learning in nonlinear inverse problems

    William Herzberg, Daniel B Rowe, Andreas Hauptmann, and Sarah J Hamilton. “Graph convolutional networks for model-based learning in nonlinear inverse problems”. In:IEEE transactions on computational imaging7 (2021), pp. 1341–1353.doi:10.1109/TCI.2021.3132190

    arXiv 2021

  77. [77]

    DiffTaichi: Differentiable programming for physical simulation

    Yuanming Hu et al. “DiffTaichi: Differentiable programming for physical simulation”. In:8th International Conference on Learning Representa- tions (ICLR 2020). 2020

    2020

  78. [78]

    Convergent bregman plug-and-play image restoration for pois- son inverse problems

    Samuel Hurault, Ulugbek Kamilov, Arthur Leclaire, and Nicolas Pa- padakis. “Convergent bregman plug-and-play image restoration for pois- son inverse problems”. In:Advances in Neural Information Processing Systems36 (2023), pp. 27251–27280

    2023

  79. [79]

    Gradient Step Denoiser for convergent Plug-and-Play

    Samuel Hurault, Arthur Leclaire, and Nicolas Papadakis. “Gradient Step Denoiser for convergent Plug-and-Play”. In:10th International Confer- ence on Learning Representations (ICLR 2022). 2022

    2022

  80. [80]

    Deep Learning Advances in Computer Vision with 3D Data: A Survey

    Anastasia Ioannidou, Elisavet Chatzilari, Spiros Nikolopoulos, and Ioan- nis Kompatsiaris. “Deep Learning Advances in Computer Vision with 3D Data: A Survey”. In:ACM Computing Surveys50.2 (2017), Article No.: 20 (38 pp.)doi:10.1145/3042064. 51

    doi:10.1145/3042064 2017

Showing first 80 references.