Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Conditional Clifford-Steerable CNNs condition the convolutional kernel on an equivariant summary of the input feature field, restoring missing angular degrees of freedom and improving PDE forecasting.

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

2026-08-04 09:37 UTC pith:6ELBNBXE

load-bearing objection Conditional kernels are a legitimate step forward for Clifford-steerable CNNs, but the completeness claim is a conjecture, not a theorem, and the empirical section lacks code and uncertainty estimates. the 4 major comments →

arxiv 2510.14007 v2 pith:6ELBNBXE submitted 2025-10-15 cs.LG cs.AI

Conditional Clifford-Steerable CNNs for PDE Modeling

B\'alint L\'aszl\'o Szarvas , Maksim Zhdanov This is my paper
classification cs.LG cs.AI
keywords Clifford algebrasteerable CNNsequivariant neural networksPDE forecastingpseudo-Euclidean groupsimplicit kernelskernel basis completenessgeometric product
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.

The paper argues that Clifford-Steerable CNNs (CSCNNs), which build equivariant convolutions from Clifford-algebra kernel networks, rest on an incomplete kernel basis: for example, in O(2,0) the vector-vector interaction collapses to a radial scalar and loses the frequency-2 angular component. The authors propose Conditional CSCNNs, which feed the kernel network an additional equivariant multivector obtained by (masked) mean-pooling the input feature field, so that geometric products with the relative position propagate angular information into all grades. They derive the steerability constraint for these input-dependent kernels, solve it via implicit parameterization, and prove the resulting convolution remains E(p,q)-equivariant. They conjecture that this conditioning yields a complete kernel basis for every O(p,q) and validate the improved expressivity empirically on Navier-Stokes, shallow-water, and Maxwell equations, where the method consistently beats standard CSCNNs and matches state-of-the-art baselines.

Core claim

On the paper's own terms, the central discovery is that the incompleteness of the implicit Clifford kernel basis is caused by the kernel network receiving only the relative position vector as its single multivector input: all operations are (weighted) geometric products, which keep angular information locked in grade 1. Conditioning the kernel on an equivariant summary of the input field—specifically a global mean-pooled multivector—gives the network a second multivector to multiply with, so that the geometric product carries angular phase into the scalar (grade-0) part and enables higher-frequency components such as the frequency-2 term in O(2,0) vector-vector interactions. The paper derive

What carries the argument

The conditional kernel network K̂: R^{p,q} × Cl(R^{p,q})^c × Cl(R^{p,q})^c → Cl(R^{p,q})^{cout×cin} together with the kernel head H that evaluates geometric products. The conditioning operator T—masked global mean pooling over a centered ball—produces an O(p,q)-equivariant multivector stack that is concatenated with the relative position as input to K̂. This is the mechanism that supplies the extra angular information; the equivariance of T (Proposition 4.1) is what makes the whole conditional convolution E(p,q)-equivariant.

Load-bearing premise

The completeness claim rests on the unproven conjecture that a single global mean-pooled multivector supplies enough independent angular information, for every O(p,q) and every grade pair, to generate every missing irrep component; the paper only demonstrates this for O(2,0) vector interactions and explicitly leaves the full proof for future work.

What would settle it

Pick a specific signature (p,q) and a specific input/output grade pair, compute the dimension of the O(p,q)-steerable kernel space from representation theory (Clebsch-Gordan decomposition), and compute the dimension of the space of kernels realizable by a conditional kernel network over all possible conditioning multivectors. If the realizable dimension is smaller, completeness is false. A concrete candidate to check first is the O(1,2) scalar-vector or bivector-vector interaction, where the non-compact group may introduce irreps not captured by multivector grades.

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

If this is right

  • Single-layer C-CSCNNs recover the full analytical steerable kernel basis for O(2,0), including the previously missing frequency-2 vector-vector component, so depth is no longer required to compensate for kernel incompleteness.
  • Equivariant PDE surrogates built from C-CSCNNs are strictly more expressive than standard CSCNN layers at negligible computational overhead, since conditioning by pooling costs almost nothing.
  • The framework extends to non-compact groups such as O(1,2) (Minkowski spacetime), where analytical kernel bases are difficult or unknown, making relativistic electrodynamics a natural testbed.
  • The paper's experiments show conditioning improves data efficiency: C-CSCNNs outperform CSCNNs and several strong baselines with only a few hundred training trajectories.
  • Because the conditioning operator is left free, the same equivariance proof covers future choices (max pooling, learnable pooling, hierarchical conditioning) without re-deriving the constraint.

Where Pith is reading between the lines

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

  • If Conjecture 4.1 holds, C-CSCNNs would provide a universal equivariant convolution for any pseudo-Euclidean group, removing the need for group-by-group analytic kernel solving; a proof would most likely need a Wigner-Eckart-type argument linking O(p,q) irreps to multivector grades.
  • The mean-field (global pooling) condition is the crudest allowed case; conditioning on spatially local or hierarchical summaries—analogous to fast multipole methods—would be a natural testable extension that could regain pointwise adaptivity while keeping equivariance.
  • The conjecture might fail for q>0 (indefinite signature) because non-compact groups have infinite-dimensional irreps and grade structure may not cover all of them; a counterexample would be an O(p,q) irrep that no conditional kernel network can produce regardless of conditioning.
  • The method suggests a general recipe for other equivariant architectures: any kernel or message function can be made input-dependent by adding an equivariant pooled summary, with the same template-matching efficiency as long as the summary is translation-invariant.

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

4 major / 5 minor

Summary. The paper introduces Conditional Clifford-Steerable CNNs (C-CSCNNs), in which the O(p,q)-steerable convolutional kernel is conditioned on a global mean-pooled multivector derived from the input field. The authors derive the conditional steerability constraint (Lemma 4.1), prove equivariance of the implicit parameterization (Lemma 4.2) and of the resulting convolution under an equivariant pooling operator (Prop 4.1), and propose a mean-field implementation via masked global mean pooling. They argue that conditioning alleviates the incompleteness of the standard CSCNN kernel basis, giving a symbolic demonstration for O(2,0) vector-vector interactions, and state as Conjecture 4.1 that the conditional kernel basis is complete. Experiments on Navier-Stokes, shallow-water, and relativistic/non-relativistic Maxwell equations show substantial improvements over the original CSCNN and competitive performance against several baselines.

Significance. The conditional conditioning mechanism is a natural and practical extension of implicit steerable kernels. The equivariance proofs are correct and the implementation overhead is negligible, which are real strengths. If the conjectured completeness were established, the paper would make a significant theoretical contribution to equivariant CNN design, especially for non-compact groups where analytic kernel bases are hard to derive. However, the central 'complete kernel basis' claim is not proven; the only direct evidence is the O(2,0) symbolic check in Appendix A.5. The empirical results are promising but currently lack error bars and code, limiting the strength of the validation. On balance, the contribution is valuable but needs reframing or a proof to justify the headline claim.

major comments (4)
  1. [4.4 / Conjecture 4.1] The title and abstract claim that C-CSCNNs have a 'complete kernel basis', but the only formal statement is Conjecture 4.1, explicitly left unproven. The supporting evidence in Appendix A.5 is a symbolic computation for O(2,0) vector-vector interactions with two hand-chosen auxiliary multivectors; it does not establish surjectivity of the conditional kernel map for other O(p,q), other grade pairs, or the non-compact O(1,2) case used in the relativistic Maxwell experiments. Since the completeness claim is the paper's headline contribution, this is a load-bearing gap. Either prove completeness for a well-defined class of groups/grade pairs, or reformulate the title and contributions in terms of improved expressivity.
  2. [4.3 / 4.4] The completeness argument assumes that the conditioning value ζ can supply arbitrary independent angular information. In the implementation, ζ is a masked global mean of the input field over the domain. No argument is given that this pooled value can realize the auxiliary multivectors needed in Appendix A.5, nor that it provides enough independent components for all grades. For many PDE initial conditions (e.g., spatially homogeneous fields), the pooled multivector may be zero or lie in a low-dimensional subspace, in which case the conditioning degenerates to the standard CSCNN. The conjecture should at least specify the class of input fields over which completeness is claimed.
  3. [5 (Figs 2–3, Table 1)] The empirical validation reports point estimates without error bars, significance tests, or repeated-seed statistics. Baseline numbers in Table 1 are taken from previous papers and may not be directly comparable. Moreover, in SWE-5 the models differ in size, and Table 1 shows CViT-L (92M) achieving 1.56% vs C-CSCNN Large at 2.94%, which is not accurately described as 'on par'. The comparison should be made more rigorous, or the claims tempered.
  4. [5.2 / Appendix B] The implemented conditioning operator is a masked global mean over a circular subset of a rectangular grid. In the continuum, this is O(p,q)-equivariant only if the domain is invariant under the group; on the bounded, rectangular domains used for the PDE benchmarks, the equivariance is approximate. Table 2 reports the equivariance error of the convolution alone on E(2), not of the full C-CSCNN on the benchmark domains or for O(1,2). The paper should quantify the equivariance error in the actual evaluation setup or state this approximation explicitly.
minor comments (5)
  1. [Appendix A.4, Code 1] The code defines r_clifford = [r, r*cos(phi), r*sin(phi), 0], i.e., a multivector with a nonzero scalar part r. The relative position should be a pure grade-1 vector [0, r*cos(phi), r*sin(phi), 0]. The conclusion still holds, but the code should be corrected.
  2. [Example 3.1] The text refers to 'Eq. 3.1', but the equation is Definition 3.1. Please fix the cross-reference.
  3. [5.2 / Implementation] The implementation is described as JAX/Flax, but no code or repository link is provided. A reproducibility statement or code release would strengthen the paper.
  4. [Appendix D] The text says 'We provide additional results ... in Table 6', but the displayed object is 'Figure 6'. The reference should be corrected.
  5. [5.3 / SWE-5] The claim that C-CSCNNs 'perform on par with leading approaches' is at odds with Table 1, where CViT-L outperforms C-CSCNN Large. Please rephrase to reflect the actual ranking.

Circularity Check

0 steps flagged

No significant circularity: the conditional equivariance derivation is independent, and the unproven completeness claim is an open-support gap, not a circular step.

full rationale

The paper's new contribution is the derivation of a steerability constraint for conditional kernels (Lemma 4.1, with proof in Appendix A.1) and the construction of conditional Clifford-steerable convolutions whose equivariance is proven (Lemma 4.2, Proposition 4.1, Appendix A.2-A.3). These derivations do not assume the target result: equivariance of the conditional convolution is shown directly from the kernel constraint, and the operator T is required to be equivariant rather than assumed to make the method work. The paper builds on prior work by the same authors (Zhdanov et al. 2023, 2024) for implicit parameterization and the kernel head, but these are cited as background constructions and are not used to smuggle in the conditional-completeness claim. The central headline claim that the kernel basis is 'complete' is explicitly stated as Conjecture 4.1 and left unproven, with only an O(2,0) example and empirical validation. This is an unsupported theoretical claim and a correctness/rigor risk, but it is not circular: no quantity is defined in terms of the predicted result, no fitted parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation. The empirical comparison against standard CSCNNs and external baselines is also independent of the unproven conjecture. Per the review rules, an unproven conjecture without a circular reduction does not raise the circularity score.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central derivations are built on standard steerable-CNN theory and prior Clifford-steerable kernel results, which are self-consistent. The main unproven ingredient is the expressivity of the conditioned kernel network; the paper itself labels it a conjecture.

axioms (5)
  • standard math Standard steerable CNN theory (Theorem 3.1, Weiler et al. 2023) that convolution is equivariant iff the kernel is steerable.
    Used as the foundation for deriving conditional kernel constraint; standard in literature.
  • standard math Implicit parameterization lemma (Lemma 3.1, Zhdanov et al. 2023) that equivariant MLPs yield steerable kernels.
    Basis of the kernel network architecture; prior proven result.
  • standard math Equivariance of the kernel head H (from Zhdanov et al. 2024).
    Required for the composition H ∘ K to satisfy Eq. 11; prior proof cited.
  • domain assumption O(p,q)-equivariance of global mean pooling with circular masking on the discretized grid.
    Argued in Section 5.2 and Appendix B, but not proven rigorously for all p,q and finite grids.
  • ad hoc to paper Expressivity of Clifford group equivariant networks to generate all required kernel basis components (Conjecture 4.1).
    The paper itself labels this a conjecture and leaves the proof for future work; it is the key unproven assumption behind the completeness claim.

pith-pipeline@v1.3.0-alltime-deepseek · 16473 in / 11495 out tokens · 90669 ms · 2026-08-04T09:37:14.151765+00:00 · methodology

0 comments
read the original abstract

We introduce Conditional Clifford-Steerable CNNs (C-CSCNNs), a unified framework that incorporates equivariance to arbitrary pseudo-Euclidean groups and significantly improves the expressivity of standard CSCNNs. We show that the kernel basis of the standard formulation is incomplete, limiting model capacity. To address this, we augment the kernels with equivariant representations of the input feature field. We derive the equivariance constraint for these input-dependent kernels and show how it can be solved efficiently via implicit parameterization. We empirically validate on multiple PDE forecasting tasks, including fluid dynamics and relativistic electrodynamics, where our method consistently outperforms standard CSCNNs and performs on par with state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2510.14007 by B\'alint L\'aszl\'o Szarvas, Maksim Zhdanov.

Figure 1
Figure 1. Figure 1: Conditional Clifford-Steerable CNNs use auxiliary information derived from the input [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MSE for the Shallow-water equa￾tions R 2 1-step forecasting task as a func￾tion of the simulations included in the train￾ing dataset. Conditional CSCNNs outper￾form all baselines, keeping their advantage even as the training trajectories increase. Note that since C-CSCNNs process multivector fields, we embed the feature fields into their corre￾sponding multivector basis elements, which is en￾abled by the n… view at source ↗
Figure 7
Figure 7. Figure 7: Scaling For the SWE-5 benchmark, we probed the scaling properties of our approach [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean squared errors for (1) Navier-Stokes R 2 , (2) Maxwell R 3 , and (3) relativistic Maxwell R 1,2 simulation tasks as a function of the simulations included in the training dataset. Conditional CSCNNs outperform all baselines, with their advantage increasing as more data is in￾cluded in the training set [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative L 2 error of conditioned CSCNNs on the shallow water equations task at different steps of the rollout trajectories. Results are shown for component u of the wind velocity field. for the default, and conditional Clifford-Steerable convolutions. Up to numerical artifacts, condi￾tional convolutions show relative errors similar to the default convolutions, providing an experimen￾tal validation of our … view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of how the receptive field of a finite, discretized kernel changes under rota￾tions. The square support causes operations to break equivariance near the corners. In constructing the Conditioned Clifford￾Steerable Kernels, we built on the architec￾ture described in Appendix A of Zhdanov et al. (2024). We form the conditioning vec￾tor by a masked spatial mean computed for each channel c and blad… view at source ↗
Figure 6
Figure 6. Figure 6: MSE versus number of samples in the NS2 and MW3 experiments for FNO and G-FNO. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Signed residuals (prediction−ground truth) of wind velocity vector component u in the one step ahead predictions for Shallow-Water Equations R 2 . 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Solver Exactness, Learned Flexibility: Equivariant Boundary-Correction Operators for Stokes Flow

    physics.flu-dyn 2026-06 unverdicted novelty 7.0

    Hybrid exact-learned equivariant operator for incompressible Stokes flow fixes the known core kernel exactly and learns only the boundary correction as a second-kind operator, achieving high accuracy, data efficiency,...

  2. Solver Exactness, Learned Flexibility: Equivariant Boundary-Correction Operators for Stokes Flow

    physics.flu-dyn 2026-06 conditional novelty 7.0

    Learning only the Stokes boundary correction on an exact free-space core yields a 5–16× more data-efficient operator than black-box DeepONet, with geometric generalization controlled by descriptor invariance and train...

Reference graph

Works this paper leans on

55 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bekkers, Maxime W

    Erik J. Bekkers, Maxime W. Lafarge, Mitko Veta, Koen A. J. Eppenhof, Josien P. W. Pluim, and Remco Duits. Roto-translation covariant convolutional networks for medical image analysis. In Medical Image Computing and Computer Assisted Intervention - MICCAI , 2018

    2018

  2. [2]

    Bekkers, Sharvaree P

    Erik J. Bekkers, Sharvaree P. Vadgama, Rob Hesselink, Putri A. van der Linden, and David W. Romero. Fast, expressive se(n) equivariant networks through weight-sharing in position-orientation space. In The Twelfth International Conference on Learning Representations, ICLR , 2024

    2024

  3. [3]

    JAX : composable transformations of P ython+ N um P y programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake Vander P las, Skye Wanderman- M ilne, and Qiao Zhang. JAX : composable transformations of P ython+ N um P y programs, 2018. URL http://github.com/jax-ml/jax

    2018

  4. [4]

    Johannes Brandstetter, Rianne van den Berg, Max Welling, and Jayesh K. Gupta. Clifford neural layers for PDE modeling. In The Eleventh International Conference on Learning Representations, ICLR , 2023

    2023

  5. [5]

    Johann Brehmer, Pim de Haan, S \" o nke Behrends, and Taco S. Cohen. Geometric algebra transformer. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, 2023

    2023

  6. [6]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovic

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Velickovic. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. CoRR, abs/2104.13478, 2021. URL https://arxiv.org/abs/2104.13478

    Pith/arXiv arXiv 2021

  7. [7]

    Carrier, L

    J. Carrier, L. Greengard, and V. Rokhlin. A fast adaptive multipole algorithm for particle simulations. SIAM Journal on Scientific and Statistical Computing, 9 0 (4): 0 669--686, 1988. doi:10.1137/0909044

    work page doi:10.1137/0909044 1988

  8. [8]

    A Program to Build E ( N )- Equivariant Steerable CNNs

    Gabriele Cesa, Leon Lang, and Maurice Weiler. A Program to Build E ( N )- Equivariant Steerable CNNs . In International Conference on Learning Representations ( ICLR ) , 2022

    2022

  9. [9]

    Group equivariant convolutional networks

    Taco Cohen and Max Welling. Group equivariant convolutional networks. In Proceedings of the 33nd International Conference on Machine Learning ( ICML ) , 2016

    2016

  10. [10]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. Steerable cnns. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net, 2017. URL https://openreview.net/forum?id=rJQKYt5ll

    2017

  11. [11]

    Filipovich and Stephen Hughes

    Matthew J. Filipovich and Stephen Hughes. Pycharge: An open-source python package for self-consistent electrodynamics simulations of lorentz oscillators and moving point charges. Comput. Phys. Commun., 274: 0 108291, 2022. doi:10.1016/J.CPC.2022.108291. URL https://doi.org/10.1016/j.cpc.2022.108291

    arXiv 2022

  12. [12]

    Worrall, Volker Fischer, and Max Welling

    Fabian Fuchs, Daniel E. Worrall, Volker Fischer, and Max Welling. Se(3)-transformers: 3d roto-translation equivariant attention networks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, 2020

    2020

  13. [13]

    Gupta and Johannes Brandstetter

    Jayesh K. Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized PDE modeling. CoRR, abs/2209.15616, 2022. doi:10.48550/ARXIV.2209.15616. URL https://doi.org/10.48550/arXiv.2209.15616

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.15616 2022

  14. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR , 2016

    2016

  15. [15]

    F lax: A neural network library and ecosystem for JAX , 2024

    Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Z ee. F lax: A neural network library and ecosystem for JAX , 2024. URL http://github.com/google/flax

    2024

  16. [16]

    Group equivariant fourier neural operators for partial differential equations

    Jacob Helwig, Xuan Zhang, Cong Fu, Jerry Kurtin, Stephan Wojtowytsch, and Shuiwang Ji. Group equivariant fourier neural operators for partial differential equations. In Proceedings of the 40th International Conference on Machine Learning, ICML'23. JMLR.org, 2023

    2023

  17. [17]

    _ flow ( PhiFlow ): Differentiable simulations for pytorch, tensorflow and jax

    Philipp Holl and Nils Thuerey. _ flow ( PhiFlow ): Differentiable simulations for pytorch, tensorflow and jax. In International Conference on Machine Learning. PMLR, 2024

    2024

  18. [18]

    Equivariance with learned canonicalization functions

    S \' e kou - Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, and Siamak Ravanbakhsh. Equivariance with learned canonicalization functions. In International Conference on Machine Learning, ICML , 2023

    2023

  19. [19]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. URL http://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  20. [20]

    Wagner, Alistair White, Sam Hatfield, Tom Kimpson, Navid C

    Milan Klöwer, Maximilian Gelbrecht, Daisuke Hotta, Justin Willmert, Simone Silvestri, Gregory L. Wagner, Alistair White, Sam Hatfield, Tom Kimpson, Navid C. Constantinou, and Chris Hill. SpeedyWeather.jl: Reinventing atmospheric general circulation models towards interactivity and extensibility . Journal of Open Source Software, 9 0 (98): 0 6323, 2024. do...

    doi:10.21105/joss.06323 2024

  21. [21]

    Modelling long range dependencies in n-d: From task-specific to a general purpose cnn

    David M Knigge, David W Romero, Albert Gu, Erik J Bekkers, Efstratios Gavves, Jakub M Tomczak, Mark Hoogendoorn, and Jan-Jakob Sonke. Modelling long range dependencies in n-d: From task-specific to a general purpose cnn. International Conference on Learning Representations, 2023

    2023

  22. [22]

    Harry Moore, Nicholas J

    Dávid Péter Kovács, J. Harry Moore, Nicholas J. Browning, Ilyes Batatia, Joshua T. Horton, Venkat Kapil, William C. Witt, Ioan-Bogdan Magdău, Daniel J. Cole, and Gábor Csányi. Mace-off23: Transferable machine learning force fields for organic molecules, 2023

    2023

  23. [23]

    A wigner-eckart theorem for group equivariant convolution kernels

    Leon Lang and Maurice Weiler. A wigner-eckart theorem for group equivariant convolution kernels. In International Conference on Learning Representations ( ICLR ) , 2021

    2021

  24. [24]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmnO

    2021

  25. [25]

    Bekkers, and Patrick Forr \' e

    Cong Liu, Sharvaree Vadgama, David Ruhe, Erik J. Bekkers, and Patrick Forr \' e . Clifford group equivariant diffusion models for 3d molecular generation. CoRR, abs/2504.15773, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021 , pp.\ 9992--10002. IEEE , 2021. doi:10.1109/ICCV48922.2021.00986. URL https:...

    arXiv 2021

  27. [27]

    Smith, Mateusz Paprocki, Ond r ej C ert\' i k, Sergey B

    Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ond r ej C ert\' i k, Sergey B. Kirpichev, Matthew Rocklin, Amit Kumar, Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rathnayake, Sean Vig, Brian E. Granger, Richard P. Muller, Francesco Bonazzi, Harsh Gupta, Shivam Vats, Fredrik Johansson, Fabian Pedregosa, Matthew J. Curry, Andy R. Terrel, S t...

    2017

  28. [28]

    Equivariant adaptation of large pretrained models

    Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, 2023

    2023

  29. [29]

    Equivariant non-linear maps for neural networks on homogeneous spaces

    Elias Nyholm, Oscar Carlsson, Maurice Weiler, and Daniel Persson. Equivariant non-linear maps for neural networks on homogeneous spaces. CoRR, abs/2504.20974, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Fengbo: a clifford neural operator pipeline for 3d pdes in computational fluid dynamics

    Alberto Pepe, Mattia Montanari, and Joan Lasenby. Fengbo: a clifford neural operator pipeline for 3d pdes in computational fluid dynamics. In The Thirteenth International Conference on Learning Representations, ICLR , 2025

    2025

  31. [31]

    Ross, and Kamyar Azizzadenesheli

    Md Ashiqur Rahman, Zachary E. Ross, and Kamyar Azizzadenesheli. U-NO: u-shaped neural operators. Trans. Mach. Learn. Res., 2023, 2023. URL https://openreview.net/forum?id=j3oQF9coJd

    2023

  32. [32]

    Romero, Anna Kuzina, Erik J

    David W. Romero, Anna Kuzina, Erik J. Bekkers, Jakub Mikolaj Tomczak, and Mark Hoogendoorn. Ckconv: Continuous kernel convolution for sequential data. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.net/forum?id=8FhxBtXSl0

    2022

  33. [33]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Nassir Navab, Joachim Hornegger, William M. Wells III, and Alejandro F. Frangi (eds.), Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceed...

    doi:10.1007/978-3-319-24574-4 2015

  34. [34]

    Clifford group equivariant neural networks

    David Ruhe, Johannes Brandstetter, and Patrick Forr \' e . Clifford group equivariant neural networks. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, 2023 a

    2023

  35. [35]

    Clifford group equivariant neural networks

    David Ruhe, Johannes Brandstetter, and Patrick Forr \'e . Clifford group equivariant neural networks. Advances in Neural Information Processing Systems, 2023 b

    2023

  36. [36]

    Gupta, Steven De Keninck, Max Welling, and Johannes Brandstetter

    David Ruhe, Jayesh K. Gupta, Steven De Keninck, Max Welling, and Johannes Brandstetter. Geometric clifford algebra networks. In International Conference on Machine Learning, ICML , 2023 c

    2023

  37. [37]

    u tt, Pieter - Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus - Robert M \

    Kristof Sch \" u tt, Pieter - Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus - Robert M \" u ller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 2017

    2017

  38. [38]

    Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria - Florina Balcan, and Hsuan - Tien Lin (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neur...

    2020

  39. [39]

    Lorentz-equivariant geometric algebra transformers for high-energy physics

    Jonas Spinner, Victor Bres \' o , Pim de Haan, Tilman Plehn, Jesse Thaler, and Johann Brehmer. Lorentz-equivariant geometric algebra transformers for high-energy physics. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, 2024

    2024

  40. [40]

    Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley

    Nathaniel Thomas, Tess E. Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds. CoRR, abs/1802.08219, 2018. URL http://arxiv.org/abs/1802.08219

    Pith/arXiv arXiv 2018

  41. [41]

    Tycho F. A. van der Ouderaa, David W. Romero, and Mark van der Wilk. Relaxing equivariance constraints with non-stationary continuous filters. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022

    2022

  42. [42]

    Octformer: Octree-based transformers for 3D point clouds

    Peng-Shuai Wang. Octformer: Octree-based transformers for 3D point clouds. ACM Transactions on Graphics (SIGGRAPH), 42 0 (4), 2023

    2023

  43. [43]

    Physics-guided deep learning for dynamical systems: A survey

    Rui Wang. Physics-guided deep learning for dynamical systems: A survey. CoRR, abs/2107.01272, 2021. URL https://arxiv.org/abs/2107.01272

    Pith/arXiv arXiv 2021

  44. [44]

    Incorporating symmetry into deep dynamics models for improved generalization

    Rui Wang, Robin Walters, and Rose Yu. Incorporating symmetry into deep dynamics models for improved generalization. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=wta_8Hx2KD

    2021

  45. [45]

    Seidman, Shyam Sankaran, Hanwen Wang, George J

    Sifan Wang, Jacob H. Seidman, Shyam Sankaran, Hanwen Wang, George J. Pappas, and Paris Perdikaris. Cvit: Continuous vision transformer for operator learning. In The Thirteenth International Conference on Learning Representations, ICLR , 2025

    2025

  46. [46]

    General e(2)-equivariant steerable cnns

    Maurice Weiler and Gabriele Cesa. General e(2)-equivariant steerable cnns. Advances in Neural Information Processing Systems, 2019

    2019

  47. [47]

    Equivariant and coordinate independent convolutional networks

    Maurice Weiler, Patrick Forr \'e , Erik Verlinde, and Max Welling. Equivariant and coordinate independent convolutional networks. A Gauge Field Theory of Neural Networks, pp.\ 110, 2023

    2023

  48. [48]

    Solving high-dimensional pdes with latent spectral models

    Haixu Wu, Tengge Hu, Huakun Luo, Jianmin Wang, and Mingsheng Long. Solving high-dimensional pdes with latent spectral models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of Proce...

    2023

  49. [49]

    Transolver: A fast transformer solver for pdes on general geometries

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries. In International Conference on Machine Learning, 2024 a

    2024

  50. [50]

    Point transformer V3: simpler, faster, stronger

    Xiaoyang Wu, Li Jiang, Peng - Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer V3: simpler, faster, stronger. In Conference on Computer Vision and Pattern Recognition ( CVPR ) , 2024 b

    2024

  51. [51]

    Funkhouser

    Fisher Yu, Vladlen Koltun, and Thomas A. Funkhouser. Dilated residual networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pp.\ 636--644. IEEE Computer Society, 2017. doi:10.1109/CVPR.2017.75. URL https://doi.org/10.1109/CVPR.2017.75

    doi:10.1109/cvpr.2017.75 2017

  52. [52]

    Implicit convolutional kernels for steerable cnns

    Maksim Zhdanov, Nico Hoffmann, and Gabriele Cesa. Implicit convolutional kernels for steerable cnns. Advances in Neural Information Processing Systems, 2023

    2023

  53. [53]

    Clifford-steerable convolutional neural networks

    Maksim Zhdanov, David Ruhe, Maurice Weiler, Ana Lucic, Johannes Brandstetter, and Patrick Forr \'e . Clifford-steerable convolutional neural networks. Forty-first International Conference on Machine Learning, ICML , 2024

    2024

  54. [54]

    Erwin: A tree-based hierarchical transformer for large-scale physical systems

    Maksim Zhdanov, Max Welling, and Jan-Willem van de Meent. Erwin: A tree-based hierarchical transformer for large-scale physical systems. In International Conference on Machine Learning ( ICML ) , 2025

    2025

  55. [55]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...