Pith. sign in

REVIEW 1 major objections 5 minor 3 cited by

Deep equivariant networks achieve universality exactly when they can separate inputs entry-wise.

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 09:19 UTC pith:XH7ORWZF

load-bearing objection Real new idea, but Definition 6's missing equivariance condition makes the central theorems false as printed; easy fix, still deserves review. the 1 major comments →

arxiv 2510.15814 v2 pith:XH7ORWZF submitted 2025-10-17 stat.ML cs.LG

On Universality of Deep Equivariant Networks

Marco Pacini , Mircea Petrache , Bruno Lepri , Shubhendu Trivedi , Robin Walters This is my paper
classification stat.ML cs.LG MSC 68T0741A30
keywords equivariant neural networksuniversalityseparation powerentry-wise separabilitydepthreadout layerspermutation representationsinvariant 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.

Universality questions for equivariant neural networks have resisted a general answer: existing results either require impractically large hidden representations or apply only to specific architectures. This paper proves that, for networks built from permutation representations with pointwise activations, universality is governed by a refined notion of separation—entry-wise separability—which tracks the distinguishability of pairs of inputs coordinate by coordinate in the output. For invariant networks, it shows that appending a fully connected readout layer secures approximation of every continuous function consistent with the network's separation relation. For equivariant networks, it proves that once entry-wise separation stabilizes with depth, or when the output layer is a convolutional filter of width one, the network can approximate every equivariant continuous function respecting that entry-wise relation. These results identify depth and readout layers as the decisive mechanism that turns separation power into approximation power.

Core claim

The central claim is that separation-constrained universality holds for deep equivariant networks once the architecture can resolve the right separation structure. The paper introduces entry-wise separability, which records, for each output coordinate, which input pairs are forced to share that coordinate's value across all networks in the class. The main theorems state: Theorem 1 shows that for invariant networks Uσ(M1,...,Md,I,L)=Cρ(V) — adding a fully connected readout after the invariant layer gives universality within the class of continuous functions respecting the network's separation relation. Theorem 2 shows that if the entry-wise separation relation of an equivariant network stabil

What carries the argument

Entry-wise separability is the central object: a vector of separation relations, one per output coordinate, defined via projections onto coordinate lines of the output space. The proofs also rely on a reconstruction map θ*_x that rebuilds a G-equivariant function from its scalar G_x-invariant projections, and on a parallelization lemma that bundles multiple invariant maps into a single wider network. The layer space C of width-one convolutional filters and the invariant layer space I with a fully connected readout L are the architectural levers that, respectively, serve as output surrogates for full readouts and provide the final approximation step via the classical universal approximation o

Load-bearing premise

The proofs assume that the target class Cρ in Theorems 2–3 and Proposition 2 is defined to contain only G-equivariant functions that respect the entry-wise separation relations; as printed, Definition 6 admits all continuous functions (including non-equivariant ones) that respect those relations, which would make the equality with the network's universality class false.

What would settle it

Take the two-node convolutional network of Example 3 with G=S2 acting on X={1,2}. The entry-wise separation relations are ρ1={(α,β):α1=β1} and ρ2={(α,β):α2=β2}. The continuous function f(x)=(x1,0) respects both ρ1 and ρ2 but is not S2-equivariant. If f belongs to the closure Uσ(C,...,C) for depth d≥2, then the theorem's equality U=Cρ fails under the printed definition; if it does not, then the equalities in the paper require the equivariant restriction of Cρ that the proofs implicitly use.

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

If this is right

  • For invariant networks, any architecture whose separation relation is known becomes universal as soon as a fully connected readout layer is appended; depth alone does not suffice unless the readout is present.
  • For equivariant networks, universality can be certified by checking that entry-wise separation stabilizes with depth; this occurs after a finite, architecture-dependent number of layers.
  • A convolutional output layer of width one yields entry-wise separation universality without waiting for separation to stabilize, meaning the output layer choice can shortcut the need for extra depth.
  • The exact class of approximable functions is the entry-wise separation-constrained class: every continuous equivariant function that respects the per-coordinate separation relations is in the closure, and nothing else is.
  • These results unify earlier specialized universality theorems for sum-pooling set networks and graph neural networks under a single separation-constrained framework.

Where Pith is reading between the lines

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

  • The entry-wise criterion suggests a practical diagnostic: measure per-output-coordinate distinguishability on finite samples; if it is not yet stable, adding depth should increase expressive range, while once stable, further depth changes only representational efficiency, not the set of approximable functions.
  • Because Theorem 3 does not require separation stabilization, it suggests that the design of equivariant architectures should place special weight on the output head: a width-one convolutional readout can confer universality even when the body of the network is not deep.
  • The results leave open the question of quantitative depth thresholds: given a specific group and layer space, how many layers are needed for separation stabilization? A concrete bound would turn the universality theorem into an engineering recipe.
  • If one reads the paper's definition of Cρ literally rather than as the equivariant restriction used in the proofs, counterexamples exist (e.g., non-equivariant coordinate projections in the CNN example); the intended reading is that Cρ consists only of G-equivariant functions, and that restriction is essential.

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

1 major / 5 minor

Summary. The paper develops a general theory of separation-constrained universality for invariant and equivariant neural networks with point-wise activations and permutation representations. It proves that invariant networks with a fully connected readout layer are universal in the class of continuous functions respecting the network's separation relation (Theorem 1). For equivariant networks, it introduces the notion of entry-wise separability (Definition 6), argues that standard separability is insufficient (Example 3), and claims two universality theorems: deep networks become entry-wise separation-universal once separation stabilizes with depth (Theorem 2 and Corollary 1), and a width-1 convolutional readout achieves the same without requiring depth stabilization (Theorem 3). The paper positions these results as unifying and extending prior architecture-specific results, including those of Segol & Lipman and Zaheer et al.

Significance. The conceptual contribution is valuable: identifying depth and readout layers as the mechanism that converts separation constraints into universality in equivariant networks, and introducing entry-wise separability as the correct refinement of standard separation. The appendix contains substantial proofs and the paper recovers several known results as special cases. However, the central equivariant statements are technically false as printed because Definition 6 omits G-equivariance from the target class C_rho, a load-bearing defect. The intended fix is clear from the appendix, which consistently works with scalar G_x-invariant projections. No code or experiments are provided; the paper is purely theoretical.

major comments (1)
  1. [Definition 6; Theorems 2, 3; Proposition 2] Definition 6 defines C_rho(V,R^X) as {f in C(V,R^X) | pi_x f respects rho_x(N) for all x} without imposing G-equivariance. Since every element of U_sigma is G-equivariant, the equalities U_sigma = C_rho in Theorems 2, 3 and Proposition 2 are impossible as printed. Concretely, in Example 3 with G=S_2, n=2, the relations rho_1,rho_2 force only x_1-dependence of the first coordinate and x_2-dependence of the second; the map (x_1,x_2)->(x_1,0) lies in the printed C_rho but is not S_2-equivariant and is not in U_sigma(C,C) by Proposition 4. The appendix proofs use the equivariant target C_{G_x,rho}(V) (Lemma 4, Eq. (10)), showing the intended restriction. Fix: add f in C_G(V,R^X) to Definition 6, or intersect C_rho with C_G throughout.
minor comments (5)
  1. [Theorem 1 proof] The proof refers to 'Equation 5' and 'Equation 6', but no equations are numbered in the typeset text; please label the displayed equations or replace with explicit equation numbers.
  2. [Theorem 1 proof] The line 'A_h = A_h = A'_h' appears to contain a typo; likely intended 'A_h = A'_h'.
  3. [Section 5.1, after Definition 6] The sentence 'N ⊆ C_{rho(N)}(V,R^X) ⊆ C_{rho(N)}(V,R^X)' uses the same symbol for the entry-wise and standard separation targets; please distinguish them, e.g., C_rho^ew and C_rho^std.
  4. [Notation] C_rho(V) is used without specifying the output space; define C_rho(V):=C_rho(V,R) once, to avoid confusion with C_rho(V,R^X).
  5. [Appendix B.2, Lemma 4] The proof of Lemma 4 asserts 'rho = rho({phi_{j,i} o sigma o theta | ...})' without derivation; a short justification that the basis property of the phi_{j,i} transfers the separation relation would improve readability.

Circularity Check

0 steps flagged

No circular derivation in the central universality theorems; self-citations concern separation stabilization, not the target result, and the printed Definition 6 equivariance gap is a correctness defect rather than circularity.

full rationale

The derivation chain for Theorem 1 is built on Lemma 3 (Stone-Weierstrass factorization through a separating family) plus a width/parallelization argument showing the architecture realizes the compositions eta o F_h. The cited Theorem 4 of Pacini et al. (2024b) is used only to note that uniform closure preserves the separation relation, an elementary fact, not the target universality. Theorems 2 and 3 reduce equivariant universality to projected invariant classes via Proposition 3 and Equation 10, then reuse the Theorem 1 mechanism; the inclusion U_sigma subset C_rho is definitionally necessary, but the opposite inclusion is proved, so the equality is not a definitional identity. The principal self-citation, Theorem 3 of Pacini et al. (2024b) used for Corollary 1, concerns stabilization of separation power in equivariant networks; it is parameter-free, concerns a different object from the entry-wise universality equality being derived, and does not assume the target conclusion. Thus no 'prediction' reduces by construction to fitted data, a definitional identity, or a self-citation chain. A real defect exists: Definition 6 defines C_rho(V,R^X) without requiring G-equivariance, so as printed Proposition 2 and Theorems 2-3 are false (e.g., (x1,x2) -> (x1,0) lies in C_rho for the S_2 example but not in U_sigma); the appendix's use of C_{G_x,rho} indicates the intended equivariant restriction. This is a serious correctness gap but not circularity. The score 2 reflects moderate dependence on prior self-cited theorems without independent machine-checked verification; no specific circular step is identified.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The theory is self-contained except for standard approximation theorems and prior separation results from the same group of authors; no free parameters or fitted values appear. The main unstated assumption is the non-polynomial activation requirement, and the main definitional issue is the missing equivariance restriction in C_ρ.

axioms (5)
  • domain assumption Finite group, permutation representations, pointwise continuous activation; layer spaces of form (1).
    Scope restrictions stated in Section 3.2; the theorems do not cover compact continuous groups or non-permutation representations.
  • standard math Universal approximation theorem (Pinkus) applies: networks with activation σ and variable width are dense in C(R^h).
    Invoked in proof of Theorem 1 (Equation 5) and Lemma 4; requires σ non-polynomial, which the paper does not state.
  • standard math Stone-Weierstrass theorem.
    Used in Lemma 3 to show the algebra generated by projections is dense in C_ρ.
  • domain assumption Separation preservation/stabilization results of Pacini et al. (2024b), Theorem 3 and Theorem 4.
    Theorems 1 and 2 use these prior self-cited results; no proof is given in this paper.
  • standard math The equivalence relation ρ is closed, so C_ρ is a closed subspace and Lemma 3 applies.
    Follows from ρ being the intersection of equalizer sets of continuous functions, but is not explicitly verified in the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 21290 in / 25149 out tokens · 199824 ms · 2026-08-04T09:19:40.936913+00:00 · methodology

0 comments
read the original abstract

Universality results for equivariant neural networks remain rare. Those that do exist typically hold only in restrictive settings: either they rely on regular or higher-order tensor representations, leading to impractically high-dimensional hidden spaces, or they target specialized architectures, often confined to the invariant setting. This work develops a more general account. For invariant networks, we establish a universality theorem under separation constraints, showing that the addition of a fully connected readout layer secures approximation within the class of separation-constrained continuous functions. For equivariant networks, where results are even scarcer, we demonstrate that standard separability notions are inadequate and introduce the sharper criterion of $\textit{entry-wise separability}$. We show that with sufficient depth or with the addition of appropriate readout layers, equivariant networks attain universality within the entry-wise separable regime. Together with prior results showing the failure of universality for shallow models, our findings identify depth and readout layers as a decisive mechanism for universality, additionally offering a unified perspective that subsumes and extends earlier specialized results.

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. On the Expressive Power of Permutation-Equivariant Weight-Space Networks

    cs.LG 2026-02 conditional novelty 7.0

    Permutation-equivariant weight-space networks are all equally expressive, and universality holds when hidden-layer biases are pairwise distinct.

  2. Data Augmentation: A Fourier Analysis Perspective

    cs.LG 2026-06 unverdicted novelty 6.0

    Partial random data augmentation matches full group augmentation's minimax rates up to vanishing approximation error for classical learning problems, but exact invariance requires the full group for expressive hypotheses.

  3. Drawback of Enforcing Equivariance and its Compensation via the Lens of Expressive Power

    cs.LG 2025-12 unverdicted novelty 6.0

    Enforcing equivariance reduces expressive power in 2-layer ReLU networks but enlarging the model compensates with proven size bounds and yields lower hypothesis space dimensionality for better generalization.

Reference graph

Works this paper leans on

41 extracted references · 11 linked inside Pith · cited by 3 Pith papers

  1. [1]

    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...

  2. [2]

    E(n) Equivariant Topological Neural Networks , February 2025

    Claudio Battiloro, Ege Karaismailoğlu, Mauricio Tec, George Dasoulas, Michelle Audirac, and Francesca Dominici. E(n) Equivariant Topological Neural Networks , February 2025. URL http://arxiv.org/abs/2405.15429. arXiv:2405.15429 [cs]

    Pith/arXiv arXiv 2025

  3. [3]

    Lorentz Group Equivariant Neural Network for Particle Physics

    Alexander Bogatskiy, Brandon Anderson, Jan Offermann, Marwah Roussi, David Miller, and Risi Kondor. Lorentz Group Equivariant Neural Network for Particle Physics . In Proceedings of the 37th International Conference on Machine Learning , pp.\ 992--1002. PMLR, November 2020. URL https://proceedings.mlr.press/v119/bogatskiy20a.html. ISSN: 2640-3498

    2020

  4. [4]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. Geometric Deep Learning : Grids , Groups , Graphs , Geodesics , and Gauges . arXiv:2104.13478 [cs, stat], May 2021. URL http://arxiv.org/abs/2104.13478. arXiv: 2104.13478

    Pith/arXiv arXiv 2021

  5. [5]

    Deep Learning Architectures : A Mathematical Approach

    Ovidiu Calin. Deep Learning Architectures : A Mathematical Approach . Springer Publishing Company, Incorporated, 1st edition, 2020. ISBN 978-3-030-36720-6

    2020

  6. [6]

    On the equivalence between graph isomorphism testing and function approximation with GNNs , May 2019

    Zhengdao Chen, Soledad Villar, Lei Chen, and Joan Bruna. On the equivalence between graph isomorphism testing and function approximation with GNNs , May 2019. URL http://arxiv.org/abs/1905.12560. arXiv:1905.12560 [cs, stat]

    Pith/arXiv arXiv 2019

  7. [7]

    Group Equivariant Convolutional Networks

    Taco Cohen and Max Welling. Group Equivariant Convolutional Networks . In Proceedings of The 33rd International Conference on Machine Learning , pp.\ 2990--2999. PMLR, June 2016. URL https://proceedings.mlr.press/v48/cohenc16.html. ISSN: 1938-7228

    2016

  8. [8]

    Nadav Dym and Steven J. Gortler. Low Dimensional Invariant Embeddings for Universal Geometric Learning , May 2022. URL http://arxiv.org/abs/2205.02956. arXiv:2205.02956 [cs, math]

    Pith/arXiv arXiv 2022

  9. [9]

    Residual Pathway Priors for Soft Equivariance Constraints

    Marc Finzi, Gregory Benton, and Andrew G Wilson. Residual Pathway Priors for Soft Equivariance Constraints . In Advances in Neural Information Processing Systems , volume 34, pp.\ 30037--30049. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper/2021/hash/fc394e9935fbd62c8aedc372464e1965-Abstract.html

    2021

  10. [10]

    SE (3)- Transformers : 3D Roto - Translation Equivariant Attention Networks

    Fabian Fuchs, Daniel Worrall, Volker Fischer, and Max Welling. SE (3)- Transformers : 3D Roto - Translation Equivariant Attention Networks . Advances in Neural Information Processing Systems, 33: 0 1970--1981, 2020. URL https://proceedings.neurips.cc/paper/2020/hash/15231a7ce4ba789d13b722cc5c955834-Abstract.html?utm_source=chatgpt.com

    1970

  11. [11]

    The expressive power of kth-order invariant graph networks, July 2020

    Floris Geerts. The expressive power of kth-order invariant graph networks, July 2020. URL http://arxiv.org/abs/2007.12035. arXiv:2007.12035 [cs, math, stat]

    Pith/arXiv arXiv 2020

  12. [12]

    EXPRESSIVENESS AND APPROXIMATION PROPERTIES OF GRAPH NEURAL NETWORKS

    Floris Geerts and Juan L Reutter. EXPRESSIVENESS AND APPROXIMATION PROPERTIES OF GRAPH NEURAL NETWORKS . pp.\ 43, 2022

    2022

  13. [13]

    Fourier Transporter : Bi - Equivariant Robotic Manipulation in 3D

    Haojie Huang, Owen Lewis Howell, Dian Wang, Xupeng Zhu, Robert Platt, and Robin Walters. Fourier Transporter : Bi - Equivariant Robotic Manipulation in 3D . October 2023. URL https://openreview.net/forum?id=UulwvAU1W0

    2023

  14. [14]

    Joshi, Cristian Bodnar, Simon V

    Chaitanya K. Joshi, Cristian Bodnar, Simon V. Mathis, Taco Cohen, and Pietro Lio. On the Expressive Power of Geometric Graph Neural Networks . International Conference of Learning Representations, 2023. URL https://openreview.net/forum?id=Rkxj1GXn9_

    2023

  15. [15]

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman...

    2021

  16. [16]

    Universal Invariant and Equivariant Graph Neural Networks

    Nicolas Keriven and Gabriel Peyré. Universal Invariant and Equivariant Graph Neural Networks . In Advances in Neural Information Processing Systems , volume 32. Curran Associates, Inc., 2019. URL https://papers.nips.cc/paper_files/paper/2019/hash/ea9268cb43f55d1d12380fb6ea5bf572-Abstract.html

    2019

  17. [17]

    On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups

    Risi Kondor and Shubhendu Trivedi. On the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact Groups . In Proceedings of the 35th International Conference on Machine Learning , pp.\ 2747--2755. PMLR, July 2018. URL https://proceedings.mlr.press/v80/kondor18a.html. ISSN: 2640-3498

    2018

  18. [18]

    Lafarge, Erik J

    Maxime W. Lafarge, Erik J. Bekkers, Josien P. W. Pluim, Remco Duits, and Mitko Veta. Roto-translation equivariant convolutional networks: Application to histopathology image analysis. Medical Image Analysis, 68: 0 101849, February 2021. ISSN 1361-8415. doi:10.1016/j.media.2020.101849. URL https://www.sciencedirect.com/science/article/pii/S1361841520302139

    arXiv 2021

  19. [19]

    Invariant and Equivariant Graph Networks

    Haggai Maron, Heli Ben-Hamu, Nadav Shamir, and Yaron Lipman. Invariant and Equivariant Graph Networks . In International Conference on Learning Representations , September 2018. URL https://openreview.net/forum?id=Syx72jC9tm

    2018

  20. [20]

    Provably Powerful Graph Networks

    Haggai Maron, Heli Ben-Hamu, Hadar Serviansky, and Yaron Lipman. Provably Powerful Graph Networks . International Conference of Learning Representations, 2019 a

    2019

  21. [21]

    On the Universality of Invariant Networks

    Haggai Maron, Ethan Fetaya, Nimrod Segol, and Yaron Lipman. On the Universality of Invariant Networks . In Proceedings of the 36th International Conference on Machine Learning , pp.\ 4363--4371. PMLR, May 2019 b . URL https://proceedings.mlr.press/v97/maron19a.html. ISSN: 2640-3498

    2019

  22. [22]

    Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe

    Christopher Morris, Martin Ritzert, Matthias Fey, William L. Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and Leman Go Neural : Higher - Order Graph Neural Networks . Proceedings of the AAAI Conference on Artificial Intelligence, 33: 0 4602--4609, July 2019. ISSN 2374-3468, 2159-5399. doi:10.1609/aaai.v33i01.33014602. URL https:...

    doi:10.1609/aaai.v33i01.33014602 2019

  23. [23]

    A Characterization Theorem for Equivariant Networks with Point -wise Activations , January 2024 a

    Marco Pacini, Xiaowen Dong, Bruno Lepri, and Gabriele Santin. A Characterization Theorem for Equivariant Networks with Point -wise Activations , January 2024 a . URL http://arxiv.org/abs/2401.09235. arXiv:2401.09235 [cs] version: 1

    Pith/arXiv arXiv 2024

  24. [24]

    Separation Power of Equivariant Neural Networks , December 2024 b

    Marco Pacini, Xiaowen Dong, Bruno Lepri, and Gabriele Santin. Separation Power of Equivariant Neural Networks , December 2024 b . URL http://arxiv.org/abs/2406.08966. arXiv:2406.08966 [cs]

    Pith/arXiv arXiv 2024

  25. [25]

    On Universality Classes of Equivariant Networks , June 2025

    Marco Pacini, Gabriele Santin, Bruno Lepri, and Shubhendu Trivedi. On Universality Classes of Equivariant Networks , June 2025. URL http://arxiv.org/abs/2506.02293. arXiv:2506.02293 [cs]

    arXiv 2025

  26. [26]

    Approximation- Generalization Trade -offs under ( Approximate ) Group Equivariance

    Mircea Petrache and Shubhendu Trivedi. Approximation- Generalization Trade -offs under ( Approximate ) Group Equivariance . Advances in Neural Information Processing Systems, 36: 0 61936--61959, December 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/c35f8e2fc6d81f195009a1d2ae5f6ae9-Abstract-Conference.html

    2023

  27. [27]

    Approximation theory of the MLP model in neural networks

    Allan Pinkus. Approximation theory of the MLP model in neural networks. Acta Numerica, 8: 0 143--195, January 1999. ISSN 1474-0508, 0962-4929. doi:10.1017/S0962492900002919. URL https://www.cambridge.org/core/journals/acta-numerica/article/abs/approximation-theory-of-the-mlp-model-in-neural-networks/18072C558C8410C4F92A82BCC8FC8CF9

    doi:10.1017/s0962492900002919 1999

  28. [28]

    Qi, Su, Hao , Mo, Kaichun , and Guibas, Leonidas J

    Charles R. Qi, Su, Hao , Mo, Kaichun , and Guibas, Leonidas J. PointNet : Deep Learning on Point Sets for 3D Classification and Segmentation . In 2017 IEEE Conference on Computer Vision and Pattern Recognition ( CVPR ) , pp.\ 77--85, Honolulu, HI, July 2017. IEEE. ISBN 978-1-5386-0457-1. doi:10.1109/CVPR.2017.16. URL http://ieeexplore.ieee.org/document/8099499/

    arXiv 2017

  29. [29]

    Universal Equivariant Multilayer Perceptrons

    Siamak Ravanbakhsh. Universal Equivariant Multilayer Perceptrons . In Proceedings of the 37th International Conference on Machine Learning , pp.\ 7996--8006. PMLR, November 2020. URL https://proceedings.mlr.press/v119/ravanbakhsh20a.html. ISSN: 2640-3498

    2020

  30. [30]

    On Universal Equivariant Set Networks , January 2020

    Nimrod Segol and Yaron Lipman. On Universal Equivariant Set Networks , January 2020. URL http://arxiv.org/abs/1910.02421. arXiv:1910.02421 [cs, stat]

    Pith/arXiv arXiv 2020

  31. [31]

    Linear Representations of Finite Groups , volume 42 of Graduate Texts in Mathematics

    Jean-Pierre Serre. Linear Representations of Finite Groups , volume 42 of Graduate Texts in Mathematics . Springer, New York, NY, 1977. ISBN 978-1-4684-9460-0 978-1-4684-9458-7. doi:10.1007/978-1-4684-9458-7. URL http://link.springer.com/10.1007/978-1-4684-9458-7

    doi:10.1007/978-1-4684-9458-7 1977

  32. [32]

    Universality of group convolutional neural networks based on ridgelet analysis on groups

    Sho Sonoda, Isao Ishikawa, and Masahiro Ikeda. Universality of group convolutional neural networks based on ridgelet analysis on groups. In Proceedings of the 36th International Conference on Neural Information Processing Systems , NIPS '22, pp.\ 38680--38694, Red Hook, NY, USA, November 2022. Curran Associates Inc. ISBN 978-1-71387-108-8

    2022

  33. [33]

    Benefits of depth in neural networks, May 2016

    Matus Telgarsky. Benefits of depth in neural networks, May 2016. URL http://arxiv.org/abs/1602.04485. arXiv:1602.04485 [cs, stat]

    Pith/arXiv arXiv 2016

  34. [34]

    Tensor field networks: Rotation - and translation-equivariant neural networks for 3D point clouds, May 2018

    Nathaniel Thomas, Tess 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, May 2018. URL http://arxiv.org/abs/1802.08219. arXiv:1802.08219 [cs]

    Pith/arXiv arXiv 2018

  35. [35]

    E(n) Equivariant Graph Neural Networks

    Victor Garcia Satorras , Emiel Hoogeboom, and Max Welling. E(n) Equivariant Graph Neural Networks . In Proceedings of the 38th International Conference on Machine Learning , pp.\ 9323--9332. PMLR, July 2021. URL https://proceedings.mlr.press/v139/satorras21a.html. ISSN: 2640-3498

    2021

  36. [36]

    Error bounds for approximations with deep ReLU networks

    Dmitry Yarotsky. Error bounds for approximations with deep ReLU networks. Neural Networks, 94: 0 103--114, October 2017. ISSN 0893-6080. doi:10.1016/j.neunet.2017.07.002. URL https://www.sciencedirect.com/science/article/pii/S0893608017301545

    doi:10.1016/j.neunet.2017.07.002 2017

  37. [37]

    Optimal approximation of continuous functions by very deep ReLU networks, June 2018

    Dmitry Yarotsky. Optimal approximation of continuous functions by very deep ReLU networks, June 2018. URL http://arxiv.org/abs/1802.03620. arXiv:1802.03620 [cs]

    Pith/arXiv arXiv 2018

  38. [38]

    Deep Sets

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep Sets . In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017. URL https://papers.nips.cc/paper_files/paper/2017/hash/f22e4747da1aa27e363d86d40ff442fe-Abstract.html

    2017

  39. [39]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  40. [40]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  41. [41]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...