Pith. sign in

REVIEW 3 major objections 5 minor 53 references

A neural operator view on U-Nets for inverse imaging problems

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that although U-shaped neural operator architectures are resolution-invariant by design, a classical U-Net is more robust to resolution changes than expected, and that this changes the practical cost-benefit of operator…

desk verdict A solid, readable survey of neural-operator U-Nets whose headline robustness claim is plausible but under-anchored by an untuned, unequal-parameter comparison. read the letter →

arxiv 2608.05839 v1 pith:36B5WKCQ submitted 2026-08-06 math.NA cs.LGcs.NA

Alexander Auras , Martin Burger , Samira Kabri , Michael Moeller , Michael Schopf-Kuester This is my paper · ORCID
classification math.NAcs.LGcs.NA MSC 65J2068T0768U10
keywords neuraloperatorsU-NetinverseproblemsresolutioninvarianceFourieroperatorlimited-angleCTimagereconstructionconvolutional
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper studies what happens to U-Net-style learned reconstructions as the discretization of an ill-posed inverse problem is refined, the limit in which the discretized problem becomes truly ill-posed. The authors reinterpret convolutional layers as discretized integral operators, deriving spectral U-Nets whose kernels are parametrized by Fourier coefficients and a differential U-Net whose layers approximate derivatives, and they test all variants on post-processing of limited-angle CT reconstructions. Their central empirical claim is that the spectral U-shaped architectures are indeed resolution-invariant and generalize cleanly to unseen resolutions, but the classical U-Net is much more robust to resolution changes than expected, especially when inputs are resized to the training resolution. A reader should care because it separates the theoretical property of resolution invariance from practical robustness: the operator U-Nets pay a large parameter and memory price, while a classical U-Net with simple resizing delivers competitive performance across a wide resolution range.

What carries the argument

The load-bearing object is the spectral convolutional layer $\hat\kappa\,\hat{*}\,u := F^{-1}(\hat\kappa \cdot Fu)$, where the convolution kernel is parametrized by its Fourier coefficients and the kernel is resized to the input by zero-padding in the frequency domain. For trigonometric polynomials, this layer is equivariant with respect to trigonometric interpolation, which is what makes the spectral U-Net resolution-invariant by design. The argument also rests on the observation that a fixed spatial kernel has receptive field $[0,m/n]^2$, which shrinks as the resolution $n$ grows, so a classical discrete U-Net has no consistent continuum interpretation unless the kernel is reinterpreted. Corollary 4 quantifies the equivalence between zero-padded spatial kernels and zero-padded spectral kernels up to scaling, and the paper uses this equivalence to transform each classical U-Net block into a spectral block. For the differential route, the key identity is $\operatorname{Diff}_n(\kappa,u) = n\,\kappa*u + (1-n)\tilde\kappa\,u$, which converges to $\tilde\kappa u + \langle b,\nabla u\rangle$ as resolution grows, so the layer can represent both a directional derivative and a multiple of the identity, but the scaling makes naive differential U-Nets unstable.

What would settle it

Retrain all six architectures on the same limited-angle CT data with per-architecture hyperparameter tuning and run the full resolution range on hardware with enough memory, then compare MSE on unseen resolutions such as 384, 448, and 512 against resized-input classical U-Nets; if the classical U-Net's robustness disappears or the spectral U-Nets dominate, the paper's central empirical conclusion fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a U-Net does not have to be a neural operator to behave nearly like one. The authors design spectral (Fourier-parametrized) U-Nets, spectral resizing U-Nets, and differential U-Nets, and compare them with a classical U-Net, CNO, and U-NO on limited-angle CT post-processing at resolutions 64, 128, and 256, including mixed-resolution training and testing on unseen resolutions up to 512. They find that the classical U-Net, trained at one resolution, generalizes to a broader range of resolutions than expected, and resizing test inputs to the training resolution drastically improves generalization, so the robustness is achieved by a discretized network with cheap resizing, not by an inherently resolution-invariant operator. The spectral U-Nets show lower run-to-run variance and clean generalization to high resolutions, but resizing does not help them, and they consume far more parameters (roughly 363 million versus 32 million) and memory. The paper therefore claims that resolution invariance is a design property of operator U-Nets, but practical resolution robustness is not exclusive to them.

Load-bearing premise

The comparison assumes that the untuned hyperparameters and memory-limited baselines are representative: the authors state they did not perform an extensive hyperparameter search, and CNO and U-NO could only be tested below 512x512 and 256x256, so better-tuned or fully-resolved baselines might change the relative robustness ranking.

Editorial extensions

If this is right

  • A classical U-Net trained on one resolution can be applied to a wide range of other resolutions by resizing the input, with performance close to that of native-resolution training in the paper's CT experiments up to 512x512.
  • Spectral U-Nets and their resizing variant generalize to unseen high resolutions without resizing and show lower variance between training runs than the classical U-Net.
  • Training on mixed resolutions helps the classical U-Net generalize to unseen low resolutions, at a slight cost on high resolutions, while spectral U-Nets gain no such benefit from mixed training.
  • A naive differential U-Net, with convolutions replaced by derivative-approximating layers, performs substantially worse and is not resolution-invariant, so resolution robustness requires additional mechanisms beyond derivative layers.
  • The high parameter counts and memory footprints of CNO and U-NO limit their practical advantage: in these experiments, CNO could only be tested below 512x512 and U-NO below 256x256.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A plausible design principle suggested by the toy example: when unrolling an iterative reconstruction into a network that must transfer across resolutions, parametrize the learned kernel that excludes the skip connection rather than the one that includes it, because interpolating the former preserves the sparsity-promoting structure.
  • If the CT finding carries over to noisier and more realistic data, practitioners may prefer a standard U-Net with input resizing over operator U-Nets, since similar robustness is achieved with roughly an order of magnitude fewer parameters.
  • The differential-layer analysis implies that a resolution-invariant differential U-Net would need an explicit mechanism for jumps, such as thresholding or scale-aware normalization, because derivative-approximating layers diverge at edges as resolution grows.
  • A testable extension would measure whether the classical U-Net's robustness persists under different resizing schemes, such as bicubic instead of bilinear interpolation, and across different inverse problems; the paper's claim is empirical and tied to bilinear resizing and ellipse-phantom CT data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies how U-Net-style architectures for inverse imaging problems behave when the discretization resolution changes. It reviews neural-operator interpretations of convolutional layers, derives a spectral (FNO-style) U-Net and a spectral-resizing variant, discusses interpolated local kernels and differential operators, and validates the ideas on a 1D deblurring toy example and on 2D limited-angle CT post-processing. The central empirical claim is that U-shaped neural operator architectures are by design resolution-invariant, yet the classical U-Net generalizes across resolutions more robustly than expected.

Significance. The paper addresses a timely and important question: whether neural-operator design principles deliver practical resolution robustness for imaging inverse problems, or whether classical U-Nets already suffice when inputs are resized. The theoretical parts are largely expository but carefully assembled: Lemmas 1–2 and Corollaries 3–4 restate standard Fourier facts in the imaging context, Lemma 5 and Corollary 6 connect scaled convolutions to differential operators, and the 1D toy example is a clear interpretability device linking projected gradient descent iterations to kernel-interpolation choices. The CT experiments are more extensive than in many operator-learning papers: 10 training runs per configuration, reported performance ranges, a baseline, parameter counts, memory measurements, and a public code link. If the empirical comparison can be placed on a more equal footing, the observation that classical U-Nets are more robust than expected would be a valuable caution for the neural-operator design narrative.

major comments (3)
  1. [5.2.3 / Fig. 8 / Table 1] The headline robustness comparison is underdetermined by the experimental setup. Section 5.2.3 explicitly states that "we did not perform an extensive search for the various hyperparameters," so the fixed-resolution reference values used to anchor questions (Q2)–(Q4) may not represent what each architecture can achieve. The comparison is also unequal: Table 1 shows 31.76M parameters for the classical U-Net versus 124.23M, 363.15M, and 541.46M for CNO, the spectral U-Nets, and U-NO, respectively, and CNO and U-NO are tested only below 512×512 and at or below 256×256, respectively, so the higher-resolution range where the U-Net's robustness is asserted lacks the pure neural-operator baselines. Before the abstract's practical conclusion can be accepted, the authors should add a matched-budget comparison (tuned hyperparameters, comparable parameter counts or explicit compute normalization, and full-range evaluation where feasible) or explicitly restrict the claim to the tested configurations.
  2. [3.1.2 / Eq. (13)] The phrase "by design resolution-invariant" overstates the theoretical guarantee. Section 3.1.2 concedes that pointwise ReLU does not preserve band-limits, so Corollary 3 holds for the spectral convolution, not for the full spectral block S_{\hat\kappa} defined in Eq. (13). The nonlinear architecture used in the experiments is therefore not exactly resolution-invariant, even in the trigonometric-interpolation sense. The abstract and Section 6 should qualify the invariance as a property of the linear spectral convolution within the blocks, with the nonlinear activation treated as an approximation.
  3. [5.2.1] The empirical evidence for the abstract's general statement about inverse imaging problems is obtained from a single benchmark: noiseless, synthetic ellipse phantoms with a fixed limited-angle range, as described in Section 5.2.1. Noise is a central aspect of ill-posed inverse problems, and the authors deliberately exclude it to simplify training. As presented, the conclusion "the classical U-Net architecture seems to be more robust with respect to resolution changes than expected" is a statement about this specific benchmark. The authors should either add a robustness experiment with noisy data or soften the general wording so that it does not claim more than the evidence supports.
minor comments (5)
  1. [3.1.3 / Corollary 4] In the second half of Corollary 4, the definition of \hat\kappa^n_k for k\in[n]^2\setminus[m]^2 repeats "\hat\kappa_k" instead of setting the out-of-band coefficients to zero; this appears to be a typo, as the surrounding text describes zero-padding in the Fourier domain.
  2. [4.1 / Example 3] In Example 3, the condition "\|i/n\|=4" should presumably read "\|i/n\|=1/4" to match the support of the jump set \partial B_{1/4}(0); please correct this typo.
  3. [5.2.2] For the CNO and U-NO baselines, the text says hyperparameters were chosen to resemble the classical U-Net or the spectral resizing U-Net, but the actual values (depth, number of channels, kernel sizes, activation settings) are not reported; providing these values would improve reproducibility of the comparison.
  4. [5.2.3] The figures report means and ranges over 10 runs but no statistical tests or error bars; given the large run-to-run variability shown for the classical U-Net in Figure 9, a simple paired or repeated-measures comparison would strengthen the claim that one architecture is more robust than another.
  5. [5.2.3] When discussing the U-NO results, the text mentions a "strong zig-zag pattern" between even and odd resolutions and attributes it to the implementation of the spectral convolution; this is plausible, but adding a brief explanation of how even/odd real tensor sizes affect the Fourier coefficients would make the discussion self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PGD-to-network derivation is exact, the spectral and differential operator results are restated from standard Fourier/analysis facts with local proofs, and the CT conclusions are empirical comparisons against external baselines.

full rationale

The paper's claimed derivation chain is self-contained. In Section 5.1, Eqs. (21)-(24) rewrite the projected gradient descent iteration for the l1-regularized deblurring problem as a ReLU network with a skip connection; this is an equivalence by algebra (g = I - tau A*A and h = -tau A*A), not a use of the conclusion as an input. The spectral U-Net is derived in Section 3 via the discrete convolution theorem (Theorem 2) and the interpolation equivariance statements (Corollaries 3 and 4). Although these are cited from the same research group's prior work [33], they are restated in the text as elementary Fourier-analytic facts, with Lemma 1 and Theorem 2 proved locally; they do not assume the paper's target claim (resolution robustness of U-Nets) and therefore are not load-bearing self-citations. The limited-angle CT study in Section 5.2 is an empirical benchmark against the classical U-Net, CNO, and U-NO implementations from external groups [31,22]; the 'classical U-Net more robust than expected' conclusion is drawn from MSE curves on hold-out data, not from any fitted quantity renamed as a prediction. The abstract's 'by design resolution-invariant' is a description of the spectral convolution's construction, not a derived result, and the actual robustness claim is tested empirically. The paper's own caveat that hyperparameters were not extensively searched (Section 5.2.3) and that CNO/U-NO were memory-limited at high resolutions weakens the generality of the empirical comparison, but that is an experimental robustness concern, not circularity. No step in the paper reduces, by construction, to its own inputs.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical entities; the central derivations rest on standard Fourier analysis, interpolation, and finite difference theory. The main hand-chosen inputs are architecture hyperparameters, not fitted physical constants. The synthetic data distribution is the key domain assumption for the empirical claims.

free parameters (2)
  • spectral U-Net kernel sizes (m_in=256, m_3=128, m_2=64, m_1=32, m_0=16) = 256, 128, 64, 32, 16
    Chosen by hand in Section 5.2.2 to match a depth-4 U-Net; affect high-frequency behavior and parameter count.
  • network hyperparameters (initial channels=64, depth=4, CNO/U-NO settings) = depth 4, 64 channels
    Not tuned; authors state this may explain U-NO underperformance in Section 5.2.3.
assumptions (3)
  • domain assumption Measurement model integrates continuous A(u) over sensor areas (Eq. 2)
    Standard model for inverse problems; used to motivate discrete ill-conditioning.
  • domain assumption Inputs are treated as band-limited trigonometric polynomials for spectral equivariance (Lemma 1, Corollary 3)
    Spectral convolution equivariance is exact only for trigonometric polynomials; ReLU breaks band-limitedness as acknowledged in Section 3.1.2.
  • ad hoc to paper CT phantoms are random ellipses, noiseless, with limited-angle sinograms of fixed angular range
    Data distribution in Section 5.2.1 is synthetic and favorable; no noise included.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A neural operator view on U-Nets for inverse imaging problems." pith.science (2026). https://pith.science/paper/36B5WKCQ

@misc{pith2026260805839,
  author       = {Pith},
  title        = {Pith review of: A neural operator view on U-Nets for inverse imaging problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36B5WKCQ}},
  note         = {Machine review of arXiv:2608.05839}
}
read the original abstract

Deep neural networks have shown great empirical success in the solution of a wide variety of ill-posed inverse problems in imaging. Yet, very few works have studied their behavior in the limit that turns the discretized ill-conditioned problems into truly ill-posed ones, i.e., for an increasing resolution of the discretization. In this work, we review common approaches to neural operator learning in architectures that resemble a U-Net, one of the most common classical architectures for inverse imaging problems. We discuss advantages and drawbacks of the respective approaches, consider a 1D toy example for improved interpretability, and present extensive numerical experiments on how different types of neural operator U-Nets can improve a first (crude) limited angle CT-reconstruction. In particular, we study how well networks trained for a certain resolution of the discretization generalize to other resolutions. Our finding is that while U-shaped neural operator architectures are by design resolution-invariant, the classical U-Net architecture seems to be more robust with respect to resolution changes than expected.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [1]

    H. W. Engl, M. Hanke, and A. Neubauer.Regularization of inverse problems. Vol. 375. Dordrecht: Kluwer, 1996

    work page 1996

  2. [2]

    Gradient Step Denoiser for convergent Plug-and-Play

    S. Hurault, A. Leclaire, and N. Papadakis. “Gradient Step Denoiser for convergent Plug-and-Play”. In:International Conference on Learning Representations. 2023

    work page 2023

  3. [3]

    Plug-and- Play priors for model based reconstruction

    S. V. Venkatakrishnan, C. A. Bouman, and B. Wohlberg. “Plug-and- Play priors for model based reconstruction”. In:2013 IEEE Global Conference on Signal and Information Processing. 2013, pp. 945–948. doi:10.1109/GlobalSIP.2013.6737048

    work page arXiv 2013

  4. [4]

    Score-based diffusion models as principled priors for inverse imaging

    B. T. Feng et al. “Score-based diffusion models as principled priors for inverse imaging”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 10520–10531

    work page 2023

  5. [5]

    Multilevel Diffusion: Infinite Dimensional Score- Based Diffusion Models for Image Generation

    P. Hagemann et al. “Multilevel Diffusion: Infinite Dimensional Score- Based Diffusion Models for Image Generation”. In:SIAM Journal on Mathematics of Data Science7.3 (2025), pp. 1337–1366.doi:10.1137/ 23M1614092. eprint:https://doi.org/10.1137/23M1614092

    doi:10.1137/23m1614092 2025

  6. [6]

    Score-Based Diffusion Models in Function Space

    J. H. Lim et al. “Score-Based Diffusion Models in Function Space”. In: Journal of Machine Learning Research26.158 (2025), pp. 1–62

    work page 2025

  7. [7]

    Generative modeling by estimating gradients of the data distribution

    Y. Song and S. Ermon. “Generative modeling by estimating gradients of the data distribution”. In:Advances in neural information processing systems32 (2019)

    work page 2019

  8. [8]

    Welker et al.Position-Blind Ptychography: Viability of image recon- struction via data-driven variational inference (in preparation)

    S. Welker et al.Position-Blind Ptychography: Viability of image recon- struction via data-driven variational inference (in preparation). 2025. 36

    work page 2025

Show all 53 references
  1. [9]

    Provably Robust Score-Based Diffusion Poste- rior Sampling for Plug-and-Play Image Reconstruction

    X. Xu and Y. Chi. “Provably Robust Score-Based Diffusion Poste- rior Sampling for Plug-and-Play Image Reconstruction”. In:Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancou- ver, BC, Canad...

    2024

  2. [10]

    Flow Matching: Markov kernels, stochastic processes and transport plans

    C. Wald and G. Steidl. “Flow Matching: Markov kernels, stochastic processes and transport plans”. In:Variational and Information Flows in Machine Learning and Optimal Transport, Oberwolfach Seminars. Vol. 56. Birkh¨ auser, 2025, pp. 185–254

    2025

  3. [11]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation

    O. Ronneberger, P. Fischer, and T. Brox. “U-Net: Convolutional Net- works for Biomedical Image Segmentation”. In:Medical Image Com- puting and Computer-Assisted Intervention – MICCAI 2015. Ed. by N. Navab et al. Cham: Springer International Publishing, 2015, pp. 234– 241

    2015

  4. [12]

    Fully Convolutional Networks for Semantic Segmentation

    E. Shelhamer, J. Long, and T. Darrell. “Fully Convolutional Networks for Semantic Segmentation”. In:IEEE Transactions on Pattern Anal- ysis and Machine Intelligence39.4 (2017), pp. 640–651.doi:10.1109/ TPAMI.2016.2572683

    2017

  5. [13]

    End-to-end Res-Unet based reconstruction algorithm for photoacoustic imaging

    J. Feng et al. “End-to-end Res-Unet based reconstruction algorithm for photoacoustic imaging”. In:Biomed. Opt. Express11.9 (Sept. 2020), pp. 5321–5340.doi:10.1364/BOE.396598

    2020 doi

  6. [14]

    Plug-and-Play Image Restoration With Deep Denoiser Prior

    K. Zhang et al. “Plug-and-Play Image Restoration With Deep Denoiser Prior”. In:IEEE Transactions on Pattern Analysis and Machine Intel- ligence44.10 (Oct. 2022), pp. 6360–6376.doi:10.1109/tpami.2021. 3088914

    2022 doi

  7. [15]

    Robustness and exploration of variational and machine learning approaches to inverse problems: An overview

    A. Auras et al. “Robustness and exploration of variational and machine learning approaches to inverse problems: An overview”. In:GAMM- Mitteilungen47.4 (2024), e202470003.doi:https : / / doi . org / 10 . 1002/gamm.202470003

    2024

  8. [16]

    Principled Approaches for Extending Neural Architec- tures to Function Spaces for Operator Learning

    J. Berner et al. “Principled Approaches for Extending Neural Architec- tures to Function Spaces for Operator Learning”. In:arXiv:2506.10973 (2025)

    2025 arXiv

  9. [17]

    Neural operator: Learning maps between function spaces with applications to pdes

    N. Kovachki et al. “Neural operator: Learning maps between function spaces with applications to pdes”. In:Journal of Machine Learning Research24.89 (2023), pp. 1–97

    2023

  10. [18]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Z. Li et al. “Fourier Neural Operator for Parametric Partial Differential Equations”. In:International Conference on Learning Representations. 2021. 37

    2021

  11. [19]

    Accelerating phase field simulations through a hybrid adaptive Fourier neural operator with U-net backbone

    C. Bonneville et al. “Accelerating phase field simulations through a hybrid adaptive Fourier neural operator with U-net backbone”. In:npj Computational Materials11.1 (Jan. 2025).doi:10 . 1038 / s41524 - 024-01488-z

    2025

  12. [20]

    Fourier neural operator with learned deformations for pdes on general geometries

    Z. Li et al. “Fourier neural operator with learned deformations for pdes on general geometries”. In:Journal of Machine Learning Research 24.388 (2023), pp. 1–26

    2023

  13. [21]

    U-shaped factorized Fourier neural operator for solving partial differential equations

    H. Liu, P. Zhao, and T. Song. “U-shaped factorized Fourier neural operator for solving partial differential equations”. In:Computers & Mathematics with Applications196 (2025), pp. 233–245.doi:https: //doi.org/10.1016/j.camwa.2025.07.013

    2025 doi

  14. [22]

    U-NO: U-shaped Neural Operators

    M. A. Rahman, Z. E. Ross, and K. Azizzadenesheli. “U-NO: U-shaped Neural Operators”. In:Transactions on Machine Learning Research (2023)

    2023

  15. [23]

    Factorized Fourier Neural Operators

    A. Tran et al. “Factorized Fourier Neural Operators”. In:The Eleventh International Conference on Learning Representations. 2023

    2023

  16. [24]

    U-FNO—An enhanced Fourier neural operator-based deep-learning model for multiphase flow

    G. Wen et al. “U-FNO—An enhanced Fourier neural operator-based deep-learning model for multiphase flow”. In:Advances in Water Re- sources163 (2022), p. 104180

    2022

  17. [25]

    Neural operators with localized integral and differential kernels

    M. Liu-Schiaffini et al. “Neural operators with localized integral and differential kernels”. In:Proceedings of the 41st International Con- ference on Machine Learning. ICML’24. Vienna, Austria: JMLR.org, 2024

    2024

  18. [26]

    Scalable and Equivariant Spherical CNNs by Discrete-Continuous (DISCO) Convolutions

    J. Ocampo, M. A. Price, and J. McEwen. “Scalable and Equivariant Spherical CNNs by Discrete-Continuous (DISCO) Convolutions”. In: The Eleventh International Conference on Learning Representations. 2023

    2023

  19. [27]

    Aubert and P

    G. Aubert and P. Kornprobst.Mathematical Problems in Image Pro- cessing: Partial Differential Equations and the Calculus of Variations. Springer New York, 2006.doi:10.1007/978-0-387-44588-5

    2006 doi

  20. [28]

    Sch¨ onlieb.Partial differential equation methods for image in- painting

    C.-B. Sch¨ onlieb.Partial differential equation methods for image in- painting. Cambridge monographs on applied and computational math- ematics 29. New York, NY: Cambridge University Press, 2015. 254 pp

    2015

  21. [29]

    Deep neural networks motivated by par- tial differential equations

    L. Ruthotto and E. Haber. “Deep neural networks motivated by par- tial differential equations”. In:Journal of Mathematical Imaging and Vision62.3 (2020), pp. 352–364

    2020

  22. [30]

    Representation equivalent neural operators: a framework for alias-free operator learning

    F. Bartolucci et al. “Representation equivalent neural operators: a framework for alias-free operator learning”. In:Advances in Neural In- formation Processing Systems36 (2023), pp. 69661–69672. 38

    2023

  23. [31]

    Convolutional neural operators for robust and ac- curate learning of PDEs

    B. Raoni´ c et al. “Convolutional neural operators for robust and ac- curate learning of PDEs”. In:Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23. New Orleans, LA, USA: Curran Associates Inc., 2023

    2023

  24. [32]

    Fourier Neural Operator for Image Classification

    W. Johnny et al. “Fourier Neural Operator for Image Classification”. In:2022 17th Iberian Conference on Information Systems and Tech- nologies (CISTI). 2022, pp. 1–6.doi:10.23919/CISTI54924.2022. 9820128

    2022

  25. [33]

    Resolution-Invariant Image Classification Based on Fourier Neural Operators

    S. Kabri et al. “Resolution-Invariant Image Classification Based on Fourier Neural Operators”. In:Scale Space and Variational Methods in Computer Vision. Springer International Publishing, 2023, pp. 236– 249.doi:10.1007/978-3-031-31975-4_18

    2023 doi

  26. [34]

    Super-Resolution Neural Operator

    M. Wei and X. Zhang. “Super-Resolution Neural Operator”. In:Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR). June 2023, pp. 18247–18256

    2023

  27. [35]

    On Universal Approxima- tion and Error Bounds for Fourier Neural Operators

    N. Kovachki, S. Lanthaler, and S. Mishra. “On Universal Approxima- tion and Error Bounds for Fourier Neural Operators”. In:Journal of Machine Learning Research22.290 (2021), pp. 1–76

    2021

  28. [36]

    Convergence Rates for Learning Linear Operators from Noisy Data

    M. V. de Hoop et al. “Convergence Rates for Learning Linear Operators from Noisy Data”. In:SIAM/ASA Journal on Uncertainty Quantifica- tion11.2 (2023), pp. 480–513.doi:10.1137/21M1442942

    2023 doi

  29. [37]

    Reinhardt, S

    N. Reinhardt, S. Wang, and J. Zech.Statistical Learning Theory for Neural Operators. 2024

    2024

  30. [38]

    Understanding the Expressivity and Trainability of Fourier Neural Operator: A Mean-Field Perspective

    T. Koshizuka et al. “Understanding the Expressivity and Trainability of Fourier Neural Operator: A Mean-Field Perspective”. In:Advances in Neural Information Processing Systems. Ed. by A. Globerson et al. Vol. 37. Curran Associates, Inc., 2024, pp. 11021–11060

    2024

  31. [39]

    Multi-level monte carlo training of neural oper- ators

    J. Rowbottom et al. “Multi-level monte carlo training of neural oper- ators”. In:arXiv preprint arXiv:2505.12940(2025)

    2025

  32. [40]

    SERNet-Former: Segmentation by Efficient-ResNet with Attention-Boosting Gates and Attention-Fusion Networks

    S. Eri¸ sen. “SERNet-Former: Segmentation by Efficient-ResNet with Attention-Boosting Gates and Attention-Fusion Networks”. In:2024 IEEE International Conference on Computer Vision and Machine In- telligence (CVMI). 2024, pp. 1–6.doi:10 . 1109 / CVMI61877 . 2024 . 10782648

    2024

  33. [41]

    Quarteroni, R

    A. Quarteroni, R. Sacco, and F. Saleri.Numerical Mathematics. Springer New York, 2007.doi:10.1007/b98885

    2007 doi

  34. [42]

    Grafakos.Classical Fourier Analysis

    L. Grafakos.Classical Fourier Analysis. Springer New York, 2014.doi: 10.1007/978-1-4939-1194-3. 39

    2014 doi

  35. [43]

    Spectral Neural Operators

    V. S. Fanaskov and I. V. Oseledets. “Spectral Neural Operators”. In: Doklady Mathematics108.S2 (Dec. 2023), S226–S232.doi:10.1134/ s1064562423701107

    2023

  36. [44]

    Tr¨ oltzsch.Optimal Control of Partial Differential Equations

    F. Tr¨ oltzsch.Optimal Control of Partial Differential Equations. Amer- ican Mathematical Society, Apr. 2010.doi:10.1090/gsm/112

    2010 doi

  37. [45]

    Deep Parametric Continuous Convolutional Neural Net- works

    S. Wang et al. “Deep Parametric Continuous Convolutional Neural Net- works”. In:Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 2018

    2018

  38. [46]

    M. S. Nixon and A. S. Aguado.Feature extraction and image processing for computer vision. 4th ed. Includes bibliographical references and index. [Place of publication not identified]: Academic Press, 2020. 1 p

    2020

  39. [47]

    J. C. Strikwerda.Finite Difference Schemes and Partial Differential Equations, Second Edition. Society for Industrial and Applied Mathe- matics, Jan. 2004.doi:10.1137/1.9780898717938

    2004 doi

  40. [48]

    Ambrosio, N

    L. Ambrosio, N. Fusco, and D. Pallara.Functions of Bounded Variation and Free Discontinuity Problems. Oxford University Press, Mar. 2000. doi:10.1093/oso/9780198502456.001.0001

  41. [49]

    H. W. Alt.Linear Functional Analysis. Springer London, 2016.doi: 10.1007/978-1-4471-7280-2

    2016 doi

  42. [50]

    Neural ordinary differential equations

    R. T. Chen et al. “Neural ordinary differential equations”. In:Advances in neural information processing systems31 (2018)

    2018

  43. [51]

    A Proposal on Machine Learning via Dynamical Systems

    W. E. “A Proposal on Machine Learning via Dynamical Systems”. In: Communications in Mathematics and Statistics5 (2017)

    2017

  44. [52]

    Stable architectures for deep neural net- works

    E. Haber and L. Ruthotto. “Stable architectures for deep neural net- works”. In:Inverse problems34.1 (2017), p. 014004

    2017

  45. [53]

    Deep limits of residual neural net- works

    M. Thorpe and Y. van Gennip. “Deep limits of residual neural net- works”. In:Research in the Mathematical Sciences10.1 (2023), p. 6. 40

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.