Pith. sign in

REVIEW 4 major objections 5 minor 109 references

Graph Lineages and Skeletal Graph Products

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper introduces graph lineages—hierarchical graph families with roughly exponential growth—and defines skeletal box and cross products whose exponential growth base is the maximum of the factor bases rather than their product, making…

desk verdict Skeletal graph products with a clean max-base cost bound; the construction is real, but the claimed right-inverse from graph sequences to graded graphs is unproved. read the letter →

arxiv 2508.00197 v1 pith:LRYHEKGG submitted 2025-07-31 cs.CV cs.LGcs.NAmath.CTmath.NA

classification cs.CVcs.LGcs.NAmath.CTmath.NA MSC 05C7618A3068R1065N55
keywords graphlineagesgradedgraphsskeletalproductscategoricalhierarchicalfamiliesmultigridconvolutionalneuralnetworksscalespaces
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

Graph lineages are hierarchical families of graphs, indexed by a level number, whose vertex and edge counts grow roughly exponentially in the level. The paper's central claim is that two new 'skeletal' product operations, skeletal box product and skeletal cross product, combine two graph lineages into a new graph lineage whose growth base is max(b1,b2) rather than the product b1b2 of the ordinary graph products. These operations are defined by universal diagrams in a category of graded graphs, so they inherit algebraic structure: the skeletal box product is associative, the skeletal cross product is commutative up to isomorphism and near-associative, and both have explicit component formulas. If correct, this gives an algebraic type theory for hierarchical model architectures that can be applied to convolutional neural networks, scale spaces, and multigrid solvers while keeping space costs logarithmic in the number of combined factors. The paper supports the claim with CNN training experiments and multigrid comparisons on boundary value problems.

What carries the argument

The load-bearing object is the graded graph: a graph $G$ equipped with a graph homomorphism $\varphi_G: G \to \hat{\mathbb{N}}$, where $\hat{\mathbb{N}}$ is the graph of nonnegative integers with successor edges and self-loops; the map assigns each vertex a level number and restricts edges to level differences $0$ or $\pm1$. A graph lineage is a graded graph whose same-level and inter-level parts both obey the exponential growth bound. The skeletal box and cross products are defined by universal double-pullback diagrams built from the templates $\hat{\mathbb{N}} \hat{\Box} \hat{\mathbb{N}}$ and $\hat{\mathbb{N}} \hat{\times} \hat{\mathbb{N}}$ with additive level numbers $l=l_1+l_2$; the products keep only edges satisfying $|\Delta l|\le 1$, encode inter-level edges through the $0/1$ sparsity structure $S$ of prolongation maps, and compute level-$L$ cardinalities by the convolution sum $\sum_{m=0}^{l} O(b_1^{m(1+\epsilon)})O(b_2^{(l-m)(1+\epsilon)})=O((l+1)\max(b_1,b_2)^{l(1+\epsilon)})$, which is absorbed into $O(\max(b_1,b_2)^{l(1+\epsilon)})$. That convolution identity carries the cost argument.

What would settle it

A concrete check: compute the vertex counts at levels L=1,...,10 for the skeletal box and cross products of two explicit lineages with bases b1<b2; if the exponential base of those counts exceeds max(b1,b2), Proposition 4 fails. Alternatively, exhibit a hierarchical graph sequence for which the minimizing prolongation matrices never attain their infimum over the allowed compact matrix manifold, which would break the right-inverse map that gives every lineage its inter-level halo.

Watch

Extended reading notes

Core claim

The central discovery is Proposition 4: the graded graph $G_1 \hat{\Box} G_2$ is again a graph lineage whenever $G_1$ and $G_2$ are, and the same holds for the skeletal cross product $G_1 \hat{\times} G_2$; the growth base of the product is $\max(b_1,b_2)$, not the product $b_1b_2$. The construction builds each level $L$ of the product from pairs of levels $(l_1,l_2)$ with $l_1+l_2=L$, keeps only edges with $|\Delta l|\le 1$, and records inter-level connectivity through Kronecker products of prolongation sparsity structures $S$. The paper proves existence and universality of these products by double-pullback diagrams (Propositions 2 and 3), shows the skeletal box product is exactly associative while the skeletal cross product is commutative and near-associative with edge-inclusion bounds, and derives unary operators (thickening, escalation to frontiers) that preserve the growth class. It then shows that a CNN built from a skeletal box-cross product of a spatial grid lineage and a feature-map lineage trains to accuracies comparable to a standard CNN, and that a recursive skeletal multigrid algorithm outperforms classical geometric multigrid on two two-dimensional boundary value problems. The authors also exhibit a continuous analog of the skeletal product in the Poincaré half-plane and define skeletal function spaces via frontiers.

Load-bearing premise

The construction assumes that optimal fine-to-coarse transfer matrices between successive levels always exist and attain their optimum, so that every growing graph family can be given the inter-level edges on which the skeletal products are built.

Editorial extensions

If this is right

  • Repeated skeletal products $\hat{\times}_{i=1}^n G_i$ and $\hat{\Box}_{i=1}^n G_i$ grow with base $\max_i b_i$ instead of $\prod_i b_i$, so combining many hierarchical spaces stays affordable.
  • The skeletal box product is exactly associative, so unparenthesized $n$-way box products are unambiguous; the skeletal cross product is commutative up to isomorphism, and its $n$-way version satisfies edge-subset bounds relative to any parenthesization.
  • A convolutional network assembled from a skeletal box-cross product of a spatial grid lineage and a feature-map lineage reaches accuracies comparable to a standard CNN on both tested image-classification benchmarks, with similar training cost.
  • A recursive multigrid solver that coarsens along one skeletal product factor at a time outperforms classical multigrid and Gauss-Seidel on the two tested boundary value problems at equal work.
  • Thickening and escalation to frontier graphs preserve the $O(b^{l(1+\epsilon)})$ growth class, so pyramid and adaptive-grid constructions remain inside the lineage formalism.

Reading between the lines

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

  • Editorial inference: the level constraint $\lceil(l_1^p+l_2^p)^{1/p}\rceil=L$ interpolates between skeletal product ($p=1$), ordinary product ($p\to\infty$), and discrete sum ($p\to0^+$), suggesting a tunable family of fractional hierarchy products that the paper mentions but does not develop.
  • Editorial inference: if the cost bound transfers to the skeletal function-space construction, functions between graph lineages could become practical compositional building blocks whose domain cost is charged per frontier volume; the paper sketches the construction but does not test it experimentally.
  • Editorial inference: the Poincaré half-plane example suggests that skeletal products of scale-space lineages converge to hyperbolic products under a logarithmic-coordinate rotation; a general continuum-limit theorem would make this precise and is not proved here.
  • Editorial inference: the recursive skeletal multigrid results imply a testable prediction—skeletal multigrid should be most effective when error modes are aligned with the factor directions of the product, and the levelwise variant should suffer when they are not, matching the paper's semi-coarsening intuition.
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

4 major / 5 minor

Summary. The paper introduces 'graph lineages' as graded graphs with exponentially growing levels and inter-level bipartite connections, and develops a category-theoretic framework in which 'skeletal' versions of the standard graph cross product and box product are defined by universal diagrams. It proves existence of the skeletal products (Propositions 2 and 3), derives component-form adjacency formulas, shows that skeletal products preserve the lineage growth bound with base equal to the maximum of the factors (Proposition 4), and analyzes near-associativity of the n-ary cross product. Additional constructions include thickening, continuous analogs, and a function-space construction, with applications to convolutional neural network architectures and multigrid solvers, supported by code and experiments.

Significance. If the central constructions are sound, the paper offers an appealing algebraic type theory for hierarchical graph architectures in which the vertex- and edge-count growth base of a product is max(b1, b2) rather than b1*b2, a genuine improvement for iterated products. The paper is commendably concrete: it gives explicit component-notation formulas, proves existence of the proposed universal objects, and ships reproducible code for the CNN and multigrid experiments. The applications demonstrate practical viability even though the empirical gains over standard baselines are modest. However, the claimed scope is wider than what is proven: the route from arbitrary graph sequences to the graded graphs used by the skeletal products is asserted rather than proved, and the function-space section is explicitly acknowledged by the authors as incomplete.

major comments (4)
  1. [§4.1] The asserted right-inverse map from graph sequences to graded graphs is not proven. The sentence 'So long as prolongation maps are optimized over compact manifolds of matrices, such as orthogonal matrices of a given structure, then inf is min and there is a right-inverse map from graph sequences to graded graphs as well' assumes that the admissible set with a prescribed sparsity pattern of O(b^l) nonzeros is compact and nonempty, but this is not established for sparse orthogonal prolongations, nor is it shown that the DR-minimizer yields a halo with a lineage-compatible growth bound. Since Section 4.1 presents this as the construction of Gd(G) from an arbitrary graph lineage, the framework's scope is narrower than claimed unless the assertion is proved or the definitions are restricted to graded graphs with given halo structures.
  2. [§5.5 (Eq. 29)] The skeletal function-space construction is explicitly incomplete. Immediately after Eq. (29) the paper states that the construction 'isn’t yet quite as algebraically self-sufficient as it looks,' and the subsequent paragraph adds edge conditions informally without giving a complete component definition, a proof that the resulting object is a graded graph or lineage, or a verification of the claimed properties. Because the abstract and conclusions present function types as one of the derived skeletal constructors, this section must either be completed or clearly labeled as an outline/conjecture rather than a derived operation.
  3. [§4.5 (after Eq. 24)] The skeletal box product is disconnected within each grade, as the paper itself notes. Eq. (23) only connects vertices with the same pair (l1, l2), so there are no intra-level edges between different decompositions of the same total level L. This is a material difference from the ordinary box product, and it weakens the asserted suitability of the skeletal box product for process-model or Laplacian-based approximation, since diffusion cannot propagate between the components of a level. The suggested remedies (thickening, two-hop truncation, box-cross union) are not analyzed for lineage preservation or cost, so the process-model claim needs either a reconnection construction or a qualification.
  4. [§4.4.2] The treatment of the n-ary skeletal cross product is only a partial algebraic characterization. Equations (14)-(19) define an unparenthesized product as an edge superset of certain parenthesizations and an alternative operator as an edge subset, but the section does not provide a universal property or a proof that these edge-inclusion inequalities hold for all parenthesizations and permutations. Since the paper advertises 'similar but not identical algebraic and category-theoretic properties,' the n-ary product deserves either a precise universal characterization or a proof of the stated inclusions, rather than an assertion.
minor comments (5)
  1. [§4.5 (Figure 10)] The text after Eq. (24) refers to Figure 10 as 'an example of a skeletal cross product,' but the figure caption and the adjacency-matrix discussion indicate that Figure 10 illustrates the skeletal box product; the caption or the text should be corrected.
  2. [Throughout] There are several typographical errors, including 'propoerties' in Section 1, 'Chararistics' in Table 2, 'Osterlee' in Section 5.3.1, and 'skeleton convolutional neural network' in the Figure 16 caption.
  3. [§2.2.1 and §4.1] The notation is confusing because the same symbol G is used both for a graph lineage and for its associated graded graph, and the phrase 'G 7→ G' is ambiguous; a distinct symbol for the graded graph, such as Gd(G), should be used consistently.
  4. [§4.6 Eq. (25)] The absorption of the factor (l+1) into the epsilon in the O(max(b1,b2)^{l(1+epsilon)}) bound is valid only for b > 1, while the definition of graph lineage allows b = 1; Proposition 4 should either exclude the b = 1 edge case or state the bound with an explicit polylogarithmic factor.
  5. [Supplemental S2] The large block-matrix displays in Sections S2.1 and S2.2 are extremely wide and difficult to read; they should be reformatted or replaced with a concise block-level description of which blocks are retained after the level-number truncation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; Proposition 4 is a direct counting argument from explicit skeletal-product definitions.

full rationale

The claimed preservation result (Proposition 4, Section 4.6) is derived, not assumed: the skeletal box/cross products are defined by explicit component formulas with vertex levels l1 + l2 (Eqs. 12–13 and 23–24), and the proof computes |phi^{-1}(l)| = sum_{m=0}^l O(b1^{m(1+epsilon)}) O(b2^{(l-m)(1+epsilon)}) = O(max(b1,b2)^{l(1+epsilon)}), with the edge count handled by a similar summation. No fitted parameter or benchmark value enters this argument. Section 2.2.1 adopts the graph-lineage terminology from the authors' prior work [87], but that self-citation supplies only the definition of the growth bound and the background DR distance measure; it is not the evidence for the product-preservation theorem. Section 4.1's asserted right-inverse from graph sequences to graded graphs via optimized prolongations is unproven, and Section 5.5 explicitly concedes that Eq. (29) "isn't yet quite as algebraically self-sufficient as it looks"; these are acknowledged limitations and assumption gaps rather than circular reductions, because the product algebra and its cost bound do not depend on those gaps being closed for arbitrary input sequences. The empirical CNN comparison in Section 5.2 benchmarks an independently implemented model against MNIST and Fashion MNIST and involves no fitted prediction that was then presented as a derived result. No specific equation or fitted parameter reduces to the paper's own inputs by construction, so the derivation chain is self-contained for the central claim.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No fitted constants; the framework is definitional. The main load-bearing assumptions are the growth-bound definition of graph lineages and the existence of optimal prolongation maps, both inherited from the authors' prior work.

assumptions (4)
  • domain assumption Graph lineages are defined by the growth bound O(b^{l^{1+epsilon}}) and by prolongation maps P optimizing a graph-graph distance DR as in [87].
    This is the definitional foundation inherited from the authors' prior work [87]; the skeletal product constructions and their cost bounds are relative to this definition.
  • standard math Graded graphs are objects of the slice category of graphs over the infinite graph N with level-preserving graph homomorphisms.
    Section 4.1 introduces this category-theoretic reformulation and uses it for the universal diagram definitions; slice categories are standard constructions.
  • ad hoc to paper The 'double pullback' diagram (Diagram 6) is a valid universal definition for the skeletal cross product, and its universality transfers from the ordinary product.
    Section 4.4.1 states this without an external reference and proves it via the transmission-of-universality diagram; if the double-pullback notion is not standard, the definition is not anchored in prior theory.
  • domain assumption Optimal prolongation matrices over compact matrix manifolds exist and give a right-inverse map from graph sequences to graded graphs.
    Section 4.1 claims the infimum is attained and the right-inverse exists; this is a non-trivial analytic assumption needed for the halo/sparsity structure of graded graphs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Lineages and Skeletal Graph Products." pith.science (2026). https://pith.science/paper/LRYHEKGG

@misc{pith2026250800197,
  author       = {Pith},
  title        = {Pith review of: Graph Lineages and Skeletal Graph Products},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LRYHEKGG}},
  note         = {Machine review of arXiv:2508.00197}
}
read the original abstract

Graphs, and sequences of growing graphs, can be used to specify the architecture of mathematical models in many fields including machine learning and computational science. Here we define structured graph "lineages" (ordered by level number) that grow in a hierarchical fashion, so that: (1) the number of graph vertices and edges increases exponentially in level number; (2) bipartite graphs connect successive levels within a graph lineage and, as in multigrid methods, can constrain matrices relating successive levels; (3) using prolongation maps within a graph lineage, process-derived distance measures between graphs at successive levels can be defined; (4) a category of "graded graphs" can be defined, and using it low-cost "skeletal" variants of standard algebraic graph operations and type constructors (cross product, box product, disjoint sum, and function types) can be derived for graded graphs and hence hierarchical graph lineages; (5) these skeletal binary operators have similar but not identical algebraic and category-theoretic properties to their standard counterparts; (6) graph lineages and their skeletal product constructors can approach continuum limit objects. Additional space-efficient unary operators on graded graphs are also derived: thickening, which creates a graph lineage of multiscale graphs, and escalation to a graph lineage of search frontiers (useful as a generalization of adaptive grids and in defining "skeletal" functions). The result is an algebraic type theory for graded graphs and (hierarchical) graph lineages. The approach is expected to be well suited to defining hierarchical model architectures - "hierarchitectures" - and local sampling, search, or optimization algorithms on them. We demonstrate such application to deep neural networks (including visual and feature scale spaces) and to multigrid numerical methods.

Figures

Figures reproduced from arXiv: 2508.00197 by the authors.

Figure 2
Figure 2. A lineage of growing “butterfly network” graphs, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. A lineage of growing grid graphs [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Definition. The thickening operator θ is extended from Nˆ to all other graded graphs G (and consistently with the foregoing definition in the special case of G = θ(Nˆ)), thus defining θ : G 7→ θ(G), by the commutative Diagram 4 in the category of graphs, universal in the θ(G) position. This definition uses θ(Nˆ) as a template for all other θ(G). Such a universal diagram is a “pullback”, indicated by the extra right-… view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Graphs Nˆ (indexed by grading number i in this illustration and in Equation 7) and θ(Nˆ) (indexed by both l and i here and in Equation 6), though now l is the grading number.) Undirected graph case is illustrated; otherwise directed edges “point” towards increasing ind…
Figure 7
Figure 7. Figure 7: A lineage of bipyramids produced by thick [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: As in Figure 5, but now for the skeletal cross product [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Paths ׈ Paths [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 13
Figure 13. Figure 13: A dilated skeletal product between two path graph lineages. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Left: a complete graph lineage, representing feature maps in a convolutional neural network. Right: the [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: The final graph in the convolutional neural net lineage from Figure 14. Note the structural similarity to [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: Comparison between a skeleton convolutional neural network, and a regular CNN with identical image [PITH_FULL_IMAGE:figures/full_fig_p024_16.png]
Figure 17
Figure 17. Figure 17: Schematic showing the entire pipeline used or the experiment in Section 5.2. The preprocessing and [PITH_FULL_IMAGE:figures/full_fig_p025_17.png]
Figure 18
Figure 18. Figure 18: Boundary problems and solutions found by a multigrid solver. For comparisons of work performed by [PITH_FULL_IMAGE:figures/full_fig_p027_18.png]
Figure 19
Figure 19. Figure 19: Comparison of our skeletal multigrid algorithm on the two boundary value problems shown in Figure 18. [PITH_FULL_IMAGE:figures/full_fig_p027_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

109 extracted references · 76 canonical work pages

  1. [1]

    E. H. A DELSON , C. H. A NDERSON , J. R. B ERGEN , P. J. B URT, AND J. M. O GDEN , Pyramid methods in image processing, RCA engineer, 29 (1984), pp. 33–41

  2. [2]

    A LVAREZ-PICALLO , D

    M. A LVAREZ-PICALLO , D. G HICA , D. S PRUNGER , AND F. ZANASI , Functorial String Diagrams for Reverse- Mode Automatic Differentiation, in 31st EACSL Annual Conference on Computer Science Logic (CSL 2023), B. Klin and E. Pimentel, eds., Leibniz International Proceedings in Informatics (LIPIcs), 252, Schloss Dagstuhl – Leibniz-Zentrum f¨ur Informatik, Dag...

  3. [3]

    C. P. S. A RAUJO , Novel neural network models for computing homothetic invariances: An image algebra notation, Journal of Mathematical Imaging and Vision, 7 (1997), pp. 69–83

  4. [4]

    A WODEY , Category theory, OUP Oxford, 2010

    S. A WODEY , Category theory, OUP Oxford, 2010

  5. [5]

    B ABAUD , A

    J. B ABAUD , A. P. WITKIN , M. BAUDIN , AND R. O. D UDA, Uniqueness of the gaussian kernel for scale-space filtering, IEEE transactions on pattern analysis and machine intelligence, (1986), pp. 26–33

  6. [6]

    B AD´IAS AND A

    A. B AD´IAS AND A. G. B ANERJEE , Neural network layer algebra: A framework to measure capacity and com- pression in deep learning, IEEE Transactions on Neural Networks and Learning Systems, 35 (2024), pp. 10380– 10393

  7. [7]

    D. H. B ALLARD , Generalizing the Hough transform to detect arbitrary shapes, Pattern recognition, 13 (1981), pp. 111–122

  8. [8]

    R. E. B ANK , T. F. D UPONT , AND H. Y SERENTANT , The hierarchical basis multigrid method , Numerische Mathematik, 52 (1988), pp. 427–458

Show all 109 references
  1. [9]

    A. G. B AYDIN , B. A. P EARLMUTTER , A. A. R ADUL , AND J. M. S ISKIND , Automatic differentiation in machine learning: a survey, Journal of Machine Learning Research, 18 (2018), pp. 1–43. http://jmlr.org/papers/v18/17-468.html

  2. [10]

    E. J. B EKKERS , B-spline cnns on lie groups, in International Conference on Learning Representations, 2020

  3. [11]

    B RADLEY , J

    T.-D. B RADLEY , J. L. G ASTALDI , AND J. TERILLA , The structure of meaning in language: parallel narratives in linear algebra and category theory, Notices of the American Mathematical Society, 71 (2023)

  4. [12]

    B RANDT , Guide to multigrid development , in Multigrid Methods: Proceedings of the Conference Held at K¨oln-Porz, November 23–27, 1981, Springer, 2006, pp

    A. B RANDT , Guide to multigrid development , in Multigrid Methods: Proceedings of the Conference Held at K¨oln-Porz, November 23–27, 1981, Springer, 2006, pp. 220–312

  5. [13]

    C ARLSSON AND R

    G. C ARLSSON AND R. B. G ABRIELSSON , Topological approaches to deep learning , in Topological Data Analysis: The Abel Symposium 2018, Springer, 2020, pp. 119–146

  6. [14]

    C HEN , C

    X. C HEN , C. G ONG , Q. WAN, L. D ENG , Y. WAN, Y. LIU, B. C HEN , AND J. L IU, Transfer learning for deep neural network-based partial differential equations solving, Advances in Aerodynamics, 3 (2021), p. 36

  7. [15]

    C HEN , B

    Y. C HEN , B. D ONG , AND J. X U, Meta-mgnet: Meta multigrid networks for solving parameterized partial differential equations, Journal of computational physics, 455 (2022), p. 110996

  8. [16]

    G. S. C RUTTWELL , B. G AVRANOVI ´C, N. G HANI , P. W ILSON , AND F. ZANASI , Categorical foundations of gradient-based learning, in European Symposium on Programming, Springer International Publishing Cham, 2022, pp. 1–28

  9. [17]

    D’A MOUR , K

    A. D’A MOUR , K. H ELLER , D. M OLDOVAN , B. A DLAM , B. A LIPANAHI , A. B EUTEL , C. C HEN , J. D EATON , J. E ISENSTEIN , M. D. H OFFMAN , ET AL ., Underspecification presents challenges for credibility in modern machine learning, Journal of Machine Learning Research, 23 (20...

  10. [18]

    D ANIELY, R

    A. D ANIELY, R. F ROSTIG , AND Y. SINGER , Toward deeper understanding of neural networks: The power of initialization and a dual view on expressivity, in Advances in Neural Information Processing Systems, D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, eds., Curran...

  11. [19]

    D ENG , J

    B. D ENG , J. Y AN, AND D. L IN, Peephole: Predicting network performance before training , arXiv preprint arXiv:1712.03351, (2017)

  12. [20]

    D ENG, The mnist database of handwritten digit images for machine learning research [best of the web] , IEEE signal processing magazine, 29 (2012), pp

    L. D ENG, The mnist database of handwritten digit images for machine learning research [best of the web] , IEEE signal processing magazine, 29 (2012), pp. 141–142

  13. [21]

    R. O. D UDA AND P. E. H ART, Use of the Hough transformation to detect lines and curves in pictures , Com- munications of the ACM, 15 (1972), pp. 11–15

  14. [22]

    E LSKEN , J

    T. E LSKEN , J. H. M ETZEN , AND F. H UTTER , Neural architecture search: A survey , Journal of Machine Learning Research, 20 (2019), pp. 1–21

  15. [23]

    F ABREGAT -HERN ´ANDEZ , J

    A. F ABREGAT -HERN ´ANDEZ , J. PALANCA , AND V. BOTTI , Exploring explainable ai: category theory insights into machine learning algorithms, Machine Learning: Science and Technology, 4 (2023), p. 045061

  16. [24]

    P. F. F ELZENSZWALB AND J. D. S CHWARTZ , Hierarchical matching of deformable shapes , in 2007 IEEE conference on computer vision and pattern recognition, IEEE, 2007, pp. 1–8

  17. [25]

    F EURER , A

    M. F EURER , A. K LEIN , K. E GGENSPERGER , J. S PRINGENBERG , M. B LUM , AND F. H UTTER , Efficient and robust automated machine learning , in Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, eds., Curran Associat...

  18. [26]

    F IEDLER , Algebraic connectivity of graphs, Czechoslovak mathematical journal, 23 (1973), pp

    M. F IEDLER , Algebraic connectivity of graphs, Czechoslovak mathematical journal, 23 (1973), pp. 298–305

  19. [27]

    F LINKOW , B

    T. F LINKOW , B. A. P EARLMUTTER , AND R. M ONAHAN , Towards correct-by-construction machine-learnt models, in 19th International Conference on Integrated Formal Methods (iFM), 2024

  20. [28]

    F OLTZ, C

    F. F OLTZ, C. L AIR , AND G. M. K ELLY, Algebraic categories with few monoidal biclosed structures or none, Journal of Pure and Applied Algebra, 17 (1980), pp. 171–177

  21. [29]

    F UKUSHIMA , Neocognitron: A hierarchical neural network capable of visual pattern recognition , Neural networks, 1 (1988), pp

    K. F UKUSHIMA , Neocognitron: A hierarchical neural network capable of visual pattern recognition , Neural networks, 1 (1988), pp. 119–130

  22. [30]

    G AVRANOVI ´C, P

    B. G AVRANOVI ´C, P. L ESSARD , A. J. D UDZIK , T. VON GLEHN , J. G. M. A RA ´UJO , AND P. VELI ˇCKOVI ´C, Position: Categorical deep learning is an algebraic theory of all architectures, in Forty-first International Con- ference on Machine Learning, 2024

  23. [31]

    G ONG , M

    S. G ONG , M. B AHRI , M. M. B RONSTEIN , AND S. Z AFEIRIOU , Geometrically principled connections in graph neural networks, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), June 2020

  24. [32]

    G RUNDMANN , V

    M. G RUNDMANN , V. K WATRA, M. H AN, AND I. E SSA, Efficient hierarchical graph-based video segmen- tation, in 2010 ieee computer society conference on computer vision and pattern recognition, IEEE, 2010, pp. 2141–2148

  25. [33]

    K. G UO, Y. H U, Y. S UN, S. Q IAN , J. G AO, AND B. Y IN, Hierarchical graph convolution network for traffic forecasting, in Proceedings of the AAAI conference on artificial intelligence, 2021, pp. 151–159

  26. [34]

    H ARTMANN , M

    D. H ARTMANN , M. M EINKE , AND W. SCHR ¨ODER , An adaptive multilevel multigrid formulation for cartesian hierarchical grid methods, Computers & Fluids, 37 (2008), pp. 1103–1125

  27. [35]

    K. H E, X. Z HANG , S. R EN, AND J. S UN, Deep residual learning for image recognition, in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. 36

  28. [36]

    H UANG , R

    R. H UANG , R. L I, AND Y. XI, Learning optimal multigrid smoothers via neural networks , SIAM Journal on Scientific Computing, 45 (2022), pp. S199–S225

  29. [37]

    I LLINGWORTH AND J

    J. I LLINGWORTH AND J. K ITTLER , The adaptive Hough transform , IEEE Transactions on Pattern Analysis and Machine Intelligence, (1987), pp. 690–698

  30. [38]

    , A survey of the Hough transform, Computer vision, graphics, and image processing, 44 (1988), pp. 87– 116

  31. [39]

    I MRICH AND S

    W. I MRICH AND S. K LAVZAR , Product graphs, structure and recognition, John Wiley & Sons, 2000

  32. [40]

    I STRATE , F

    R. I STRATE , F. S CHEIDEGGER , G. M ARIANI , D. N IKOLOPOULOS , C. B EKAS , AND A. C. I. M ALOSSI , Tapas: Train-less accuracy predictor for architecture search, in Proceedings of the AAAI conference on artifi- cial intelligence, 2019, pp. 3927–3934

  33. [41]

    E. C. J ACKSON , J. A. H UGHES , M. D ALEY, AND M. W INTER , An algebraic generalization for graph and tensor-based neural networks, in 2017 IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), IEEE, 2017, pp. 1–8

  34. [42]

    J ANSSON AND T

    Y. J ANSSON AND T. L INDEBERG , Scale-invariant scale-channel networks: Deep networks that generalise to previously unseen scales, Journal of Mathematical Imaging and Vision, 64 (2022), pp. 506–536

  35. [43]

    J IN AND S

    Y. J IN AND S. G EMAN , Context and hierarchy in a probabilistic image model, in 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), IEEE, 2006, pp. 2145–2152

  36. [44]

    J OYCE AND J

    J. J OYCE AND J. V ERSCHELDE , Algebraic representations for faster predictions in convolutional neural net- works, in International Workshop on Computer Algebra in Scientific Computing, Springer, 2024, pp. 161–177

  37. [45]

    K ALOGEROPOULOS , G

    I. K ALOGEROPOULOS , G. B OURITSAS , AND Y. PANAGAKIS , Scale equivariant graph metanetworks, in The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  38. [46]

    K ANG , D

    D. K ANG , D. R AGHAVAN, P. BAILIS , AND M. Z AHARIA , Model assertions for monitoring and improving ml models, Proceedings of Machine Learning and Systems, 2 (2020), pp. 481–496

  39. [47]

    T.-W. K E, M. M AIRE , AND S. X. Y U, Multigrid neural architectures, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 6665–6673

  40. [48]

    D. P. K INGMA AND J. B A, Adam: A method for stochastic optimization , arXiv preprint arXiv:1412.6980, (2014)

  41. [49]

    K IRSCH , J

    L. K IRSCH , J. K UNZE , AND D. B ARBER , Modular networks: Learning to decompose neural computation , in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, eds., Curran Associates, Inc., 2018

  42. [50]

    K NAUER AND K

    U. K NAUER AND K. K NAUER , Algebraic Graph Theory, Walter de Gruyter GmbH, 2019

  43. [51]

    K RIZHEVSKY , I

    A. K RIZHEVSKY , I. S UTSKEVER , AND G. E. H INTON , Imagenet classification with deep convolutional neu- ral networks, in Advances in Neural Information Processing Systems, F. Pereira, C. Burges, L. Bottou, and K. Weinberger, eds., Curran Associates, Inc., 2012

  44. [52]

    L ECUN, Y

    Y. L ECUN, Y. BENGIO , AND G. H INTON , Deep learning, nature, 521 (2015), pp. 436–444

  45. [53]

    L EHNER AND T

    C. L EHNER AND T. W ETTIG , Gauge-equivariant neural networks as preconditioners in lattice qcd , Physical Review D, 108 (2023), p. 034503

  46. [54]

    T. L EI, W. J IN, R. B ARZILAY , AND T. JAAKKOLA , Deriving neural architectures from sequence and graph kernels, in Proceedings of the 34th International Conference on Machine Learning, D. Precup and Y . W. Teh, eds., Proceedings of Machine Learning Research, 70, PMLR, 06–11 ...

  47. [55]

    J. L I, Y. RONG , H. C HENG , H. M ENG , W. HUANG , AND J. H UANG , Semi-supervised graph classification: A hierarchical graph perspective, in The World Wide Web Conference, 2019, pp. 972–982

  48. [56]

    L IANG , S

    S. L IANG , S. W. J IANG , J. H ARLIM , AND H. Y ANG, Solving pdes on unknown manifolds with machine learning, Applied and Computational Harmonic Analysis, 71 (2024), p. 101652

  49. [57]

    D. L IM, H. M ARON , M. T. L AW, J. L ORRAINE , AND J. L UCAS , Graph metanetworks for processing diverse neural architectures, in The Twelfth International Conference on Learning Representations, 2023

  50. [58]

    L INDEBERG , Scale-space theory in computer vision, Springer Science & Business Media, 2013

    T. L INDEBERG , Scale-space theory in computer vision, Springer Science & Business Media, 2013

  51. [59]

    C. L IU, B. Z OPH , M. N EUMANN , J. S HLENS , W. H UA, L.-J. L I, L. F EI-FEI, A. Y UILLE , J. H UANG , AND K. M URPHY , Progressive neural architecture search, in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 19–34

  52. [60]

    S. L IU, L. G ILES , AND A. O RORBIA , Learning a hierarchical latent-variable model of 3d shapes , in 2018 international conference on 3D vision (3DV), IEEE, 2018, pp. 542–551

  53. [61]

    Y. L IU, C. P ONCE , S. L. B RUNTON , AND J. N. K UTZ, Multiresolution convolutional autoencoders, Journal of Computational Physics, 474 (2023), p. 111801

  54. [62]

    L OV ´ASZ, Large networks and graph limits, volume 60 of american mathematical society colloquium publi- cations, American Mathematical Society, Providence, RI, 22 (2012)

    L. L OV ´ASZ, Large networks and graph limits, volume 60 of american mathematical society colloquium publi- cations, American Mathematical Society, Providence, RI, 22 (2012)

  55. [63]

    I. L UZ, M. G ALUN , H. M ARON , R. BASRI , AND I. YAVNEH, Learning algebraic multigrid using graph neural networks, in International Conference on Machine Learning, PMLR, 2020, pp. 6489–6499

  56. [64]

    M ADAN , T

    S. M ADAN , T. H ENRY, J. D OZIER , H. H O, N. B HANDARI , T. S ASAKI , F. D URAND , H. P FISTER , AND X. B OIX, When and how convolutional neural networks generalize to out-of-distribution category–viewpoint combinations, Nature Machine Intelligence, 4 (2022), pp. 146–153

  57. [65]

    M ARGENSTERN , An application of grossone to the study of a family of tilings of the hyperbolic plane , Applied Mathematics and Computation, 218 (2012), pp

    M. M ARGENSTERN , An application of grossone to the study of a family of tilings of the hyperbolic plane , Applied Mathematics and Computation, 218 (2012), pp. 8005–8018. https://www.sciencedirect.com/science/article/pii/S0096300311005698, https://doi.org/10.1016/j.amc.2011.04.014

  58. [66]

    M ARON , H

    H. M ARON , H. B EN-H AMU , N. S HAMIR , AND Y. L IPMAN , Invariant and equivariant graph networks , in International Conference on Learning Representations, 2019

  59. [67]

    M CGREIVY AND A

    N. M CGREIVY AND A. HAKIM , Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations, Nature machine intelligence, 6 (2024), pp. 1256–1269

  60. [68]

    M ELLOR , J

    J. M ELLOR , J. T URNER , A. S TORKEY , AND E. J. C ROWLEY , Neural architecture search without training, in International conference on machine learning, PMLR, 2021, pp. 7588–7598

  61. [69]

    M EMIN AND P

    E. M EMIN AND P. P EREZ , A multigrid approach for hierarchical motion estimation , in Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), IEEE, 1998, pp. 933–938

  62. [70]

    M JOLSNESS , G

    E. M JOLSNESS , G. G INDI , AND P. ANANDAN , Neural networks for model matching and perceptual organi- zation, Advances in Neural Information Processing Systems, 1 (1988)

  63. [71]

    E. D. M JOLSNESS , Neural Networks, Pattern Recognition, and Fingerprint Hallucination, Ph.D. thesis, Cali- fornia Institute of Technology, 1986

  64. [72]

    https://ncatlab.org/nlab/show/funny+tensor+ product, May 2025

    NLAB AUTHORS , funny tensor product . https://ncatlab.org/nlab/show/funny+tensor+ product, May 2025. Revision 12

  65. [73]

    O LAH AND S

    C. O LAH AND S. C ARTER , Research debt, Distill, 2 (2017), p. e5. 38

  66. [74]

    C. W. O OSTERLEE AND P. W ESSELING , On the robustness of a multiple semi-coarsened grid method , Zeitschrift Fur Angewandte Mathematik Und Mechanik, 75 (1995), pp. 251–251

  67. [75]

    P AREKH , J

    R. P AREKH , J. Y ANG , AND V. HONAVAR, Constructive neural-network learning algorithms for pattern clas- sification, IEEE Transactions on Neural Networks, 11 (2000), pp. 436–451

  68. [76]

    P ASZKE , S

    A. P ASZKE , S. G ROSS , F. M ASSA , A. L ERER , J. B RADBURY , G. C HANAN , T. K ILLEEN , Z. L IN, N. G IMELSHEIN , L. A NTIGA , ET AL ., Pytorch: An imperative style, high-performance deep learning library, in Advances in Neural Information Processing Systems 32, 2019, pp. 8024–8035

  69. [77]

    P ERERA AND V

    R. P ERERA AND V. AGRAWAL, Multiscale graph neural networks with adaptive mesh refinement for accelerat- ing mesh-based simulations, Computer Methods in Applied Mechanics and Engineering, 429 (2024), p. 117152

  70. [78]

    P ERRET , G

    B. P ERRET , G. C HIERCHIA , J. C OUSTY , S. J. F. G UIMARAES , Y. K ENMOCHI , AND L. N AJMAN , Higra: Hierarchical graph analysis, SoftwareX, 10 (2019), p. 100335

  71. [79]

    P HAM , M

    H. P HAM , M. G UAN, B. Z OPH , Q. L E, AND J. D EAN, Efficient neural architecture search via parameters sharing, in International conference on machine learning, PMLR, 2018, pp. 4095–4104

  72. [80]

    P INEAU , P

    J. P INEAU , P. V INCENT -L AMARRE , K. S INHA , V. L ARIVI `ERE , A. B EYGELZIMER , F. D’A LCH ´E BUC, E. F OX, AND H. L AROCHELLE , Improving reproducibility in machine learning research (a report from the neurips 2019 reproducibility program), Journal of machine learning re...

  73. [81]

    R AFF, A step toward quantifying independently reproducible machine learning research, Advances in Neural Information Processing Systems, 32 (2019)

    E. R AFF, A step toward quantifying independently reproducible machine learning research, Advances in Neural Information Processing Systems, 32 (2019)

  74. [82]

    R EED , A

    S. R EED , A. O ORD , N. K ALCHBRENNER , S. G. C OLMENAREJO , Z. W ANG , Y. C HEN , D. B ELOV, AND N. F REITAS , Parallel multiscale autoregressive density estimation , in International conference on machine learning, PMLR, 2017, pp. 2912–2921

  75. [83]

    R ITTER , D

    S. R ITTER , D. G. T. B ARRETT , A. S ANTORO , AND M. M. B OTVINICK , Cognitive psychology for deep neural networks: A shape bias case study , in Proceedings of the 34th International Conference on Machine Learning, D. Precup and Y . W. Teh, eds., Proceedings of Machine Learni...

  76. [84]

    R ONNEBERGER , P

    O. R ONNEBERGER , P. FISCHER , AND T. BROX, U-net: Convolutional networks for biomedical image segmen- tation, in International Conference on Medical image computing and computer-assisted intervention, Springer, 2015, pp. 234–241

  77. [85]

    S ALEHIN , M

    I. S ALEHIN , M. S. I SLAM , P. S AHA , S. N OMAN , A. T UNI , M. M. H ASAN , AND M. A. B ATEN, Automl: A systematic review on automated machine learning with neural architecture search, Journal of Information and Intelligence, 2 (2024), pp. 52–81

  78. [86]

    C. B. S COTT AND E. M JOLSNESS , Graph prolongation convolutional networks: explicitly multiscale machine learning on graphs with applications to modeling of cytoskeleton, Machine Learning: Science and Technology, 2 (2020), p. 015009

  79. [87]

    C. B. S COTT AND E. M JOLSNESS , Graph Diffusion Distance: Properties and Efficient Computation , PLoS ONE, 16 (2021), p. e0249624

  80. [88]

    S PORRING , M

    J. S PORRING , M. N IELSEN , L. F LORACK , AND P. JOHANSEN , Gaussian scale-space theory, Springer Science & Business Media, 2013

  81. [89]

    S RIVASTAVA, B

    M. S RIVASTAVA, B. N USHI , E. K AMAR , S. S HAH , AND E. H ORVITZ , An empirical analysis of backward compatibility in machine learning systems, in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020, pp. 3272–3280. 39

  82. [90]

    S T ¨UBEN , A review of algebraic multigrid, Numerical Analysis: Historical Developments in the 20th Century, (2001), pp

    K. S T ¨UBEN , A review of algebraic multigrid, Numerical Analysis: Historical Developments in the 20th Century, (2001), pp. 331–359

  83. [91]

    T IAN , Y

    K. T IAN , Y. J IANG , Z. Y UAN, B. P ENG , AND L. W ANG, Visual autoregressive modeling: Scalable image generation via next-scale prediction, Advances in neural information processing systems, 37 (2024), pp. 84839– 84865

  84. [92]

    T ROTTENBERG , C

    U. T ROTTENBERG , C. W. O OSTERLEE , AND A. S CHULLER , Multigrid methods, Academic press, 2001

  85. [93]

    V OULODIMOS , N

    A. V OULODIMOS , N. D OULAMIS , A. D OULAMIS , AND E. P ROTOPAPADAKIS , Deep learning for computer vision: A brief review, Computational intelligence and neuroscience, 2018 (2018), p. 7068349

  86. [94]

    W ATANABE , Algebraic analysis for nonidentifiable learning machines , Neural Computation, 13 (2001), pp

    S. W ATANABE , Algebraic analysis for nonidentifiable learning machines , Neural Computation, 13 (2001), pp. 899–933

  87. [95]

    W EBER , Free products of higher operad algebras, Theory and applications of categories, 28 (2013), pp

    M. W EBER , Free products of higher operad algebras, Theory and applications of categories, 28 (2013), pp. 24– 65

  88. [96]

    W ESSELING , Introduction to multigrid methods, Tech

    P. W ESSELING , Introduction to multigrid methods, Tech. Report, Institute for Computer Applications in Sci- ence and Engineering, 1995

  89. [97]

    W ORRALL AND M

    D. W ORRALL AND M. W ELLING , Deep scale-spaces: Equivariance over scale, Advances in Neural Informa- tion Processing Systems, 32 (2019)

  90. [98]

    X IAO, K

    H. X IAO, K. R ASUL , AND R. V OLLGRAF , Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, arXiv preprint arXiv:1708.07747, (2017)

  91. [99]

    X U AND Y

    T. X U AND Y. M ARUYAMA , Neural string diagrams: a universal modelling language for categorical deep learning, in Artificial General Intelligence: 14th International Conference, AGI 2021, Palo Alto, CA, USA, October 15–18, 2021, Proceedings 14, Springer, 2022, pp. 306–315

  92. [100]

    Y ANG , T

    J. Y ANG , T. D ZANIC , B. P ETERSEN , J. K UDO , K. M ITTAL , V. TOMOV, J.-S. C AMIER , T. Z HAO, H. Z HA, T. KOLEV, ET AL ., Reinforcement learning for adaptive mesh refinement, in International conference on artifi- cial intelligence and statistics, PMLR, 2023, pp. 5997–6014

  93. [101]

    Y ANG , Y

    Z. Y ANG , Y. D ONG , X. D ENG , AND L. Z HANG , Amgnet: multi-scale graph neural networks for flow field prediction, Connection Science, 34 (2022), pp. 2500–2519

  94. [102]

    Y EHUDAI , E

    G. Y EHUDAI , E. F ETAYA, E. M EIROM , G. C HECHIK , AND H. M ARON , From local structures to size general- ization in graph neural networks, in International Conference on Machine Learning, PMLR, 2021, pp. 11975– 11986

  95. [103]

    Y ING , J

    Z. Y ING , J. Y OU, C. M ORRIS , X. R EN, W. H AMILTON , AND J. L ESKOVEC , Hierarchical graph representa- tion learning with differentiable pooling, Advances in neural information processing systems, 31 (2018)

  96. [104]

    J. Y OU, J. L ESKOVEC , K. H E, AND S. X IE, Graph structure of neural networks, in Proceedings of the 37th International Conference on Machine Learning, H. D. III and A. Singh, eds., Proceedings of Machine Learning Research, 119, PMLR, 13–18 Jul 2020, pp. 10881–10891

  97. [105]

    Z HANG , A

    E. Z HANG , A. K AHANA , A. K OPANI ˇC ´AKOV ´A, E. T URKEL , R. R ANADE , J. P ATHAK , AND G. E. K ARNI - ADAKIS , Blending neural operators and relaxation methods in pde numerical solvers, Nature Machine Intelli- gence, 6 (2024), pp. 1303–1313

  98. [106]

    Z HANG , Q

    Z. Z HANG , Q. L IU, Q. H U, AND C.-K. L EE, Hierarchical graph transformer with adaptive node sampling , Advances in Neural Information Processing Systems, 35 (2022), pp. 21171–21183. 40

  99. [107]

    W. Z HU, Q. Q IU, R. C ALDERBANK , G. S APIRO , AND X. C HENG , Scaling-translation-equivariant networks with decomposed convolutional filters, Journal of Machine Learning Research, 23 (2022), pp. 1–45. http://jmlr.org/papers/v23/20-099.html

  100. [108]

    Z OPH AND Q

    B. Z OPH AND Q. L E, Neural architecture search with reinforcement learning, in International Conference on Learning Representations, 2016

  101. [109]

    Z WEIG AND J

    A. Z WEIG AND J. B RUNA, A functional perspective on learning symmetric functions with neural networks, in International Conference on Machine Learning, PMLR, 2021, pp. 13023–13032. 41

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.