Pith. sign in

REVIEW 3 major objections 5 minor 76 references

A Set-to-Set Distance Measure in Hyperbolic Space

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proposes HS2SD, a hyperbolic set-to-set distance that combines geodesic and topological terms, and reports gains over point-to-point baselines in entity matching and few-shot classification.

desk verdict A plausible hyperbolic set-to-set distance with strong few-shot and entity-matching results, but the paper's central theoretical claim about Thue-Morse bases does not hold up under scrutiny. read the letter →

arxiv 2506.18529 v1 pith:5QZMNW4X submitted 2025-06-23 cs.CV cs.LG

Pengxiang Li , Wei Wu , Zhi Gao , Xiaomeng Fan , Peilin Yu , Yuwei Wu , Zhipeng Lu , Yunde Jia
show 1 more author
Mehrtash Harandi
This is my paper
classification cs.CVcs.LG MSC 51M1005C5005C60
keywords hyperbolicgeometryset-to-setdistanceThue-MorsesequenceEinsteinmidpointtopologicalfew-shotlearningentitymatchingmetric
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

The paper proposes HS2SD, a way to measure how dissimilar two sets of points are when those points live in hyperbolic space. The distance combines a global component, the hyperbolic geodesic distance between the two sets' Einstein midpoints, with a local component, the average Frobenius norm between Thue-Morse matrix sequences built from each set's adjacency and degree matrices. The paper's claim is that the local component approximates the topological structure of a set, so the combined measure captures both where the set sits and how its members relate to one another. In experiments on entity matching, standard image classification, and few-shot classification, HS2SD improves on point-to-point hyperbolic baselines, with large gains in few-shot settings.

What carries the argument

The central object is the combined distance of Eq. (6): $d_c(S_x,S_y)=\lambda d_g(S_x,S_y)+(1-\lambda)d_t(S_x,S_y)$, where $d_g$ is the hyperbolic $\ell^1$ distance between Einstein midpoints (Eq. (9)) and $d_t$ is the average Frobenius norm of differences between Thue-Morse sequences (Eq. (8)). The Thue-Morse sequence $t_0=D$, $t_1=DA$, $t_2=DAAD$, $t_3=DAADADDA$ is the mechanism that carries the topological term: the paper uses McKay's spectral criterion to argue that polynomial words in adjacency matrix $A$ and degree matrix $D$ determine tree structure, and uses the overlap-free property of Thue-Morse words to justify stopping at four terms. The $\lambda$ adapter $G$ computes a per-pair weight from cross-attention between the two sets to balance the two terms.

What would settle it

Construct two non-isomorphic trees on the same number of vertices whose first four Thue-Morse sequences $(t_0,t_1,t_2,t_3)$ have the same pairwise Frobenius distances, or whose topological distance $d_t$ is zero; such a pair would show the four-word sequence is not complete enough to distinguish tree topologies, contradicting Corollary 4.5. The check is concrete: enumerate small trees, compute their degree and adjacency matrices, and compare the four-term Frobenius vector.

Watch

Extended reading notes

Core claim

The central claim is that the distance $d_c(S_x,S_y)=\lambda d_g(S_x,S_y)+(1-\lambda)d_t(S_x,S_y)$ with $d_g$ from Einstein midpoints and $d_t$ from Thue-Morse words in adjacency and degree matrices is a better dissimilarity for hyperbolic sets than point-to-point geodesic distance. The paper states that the first four Thue-Morse words $t_0=D$, $t_1=DA$, $t_2=DAAD$, $t_3=DAADADDA$ give a robust finite approximation of the infinite polynomial combinations that determine tree isomorphism, and that the Frobenius average in Eq. (8) therefore measures topological difference. Empirically, the paper reports that both terms and the learned $\lambda$ contribute, and that the full HS2SD beats existing hyperbolic and Euclidean baselines on the tasks tested.

Load-bearing premise

The load-bearing premise is that the first four Thue-Morse words give a complete enough basis for the algebra of adjacency and degree matrices so that Eq. (8) captures a set's topology, a completeness that is assumed rather than proved.

Editorial extensions

If this is right

  • On the few-shot benchmarks, HS2SD reports gains over Hyp-ProtoNet of 4.90 and 4.04 points on miniImagenet and 12.22 and 12.05 points on tieredImagenet for 1-shot and 5-shot, respectively, which the paper attributes to the local topological term.
  • Ablations on the three few-shot datasets show that removing $d_g$, $d_t$, or the $\lambda$ adapter each lowers accuracy, so the paper's account is that all three components contribute to the final distance.
  • On the seven entity-matching datasets, HS2SD raises the average F1 from 91.7 (CampER) to 93.7, with the largest single-dataset gains on Amazon-Google (+5.8) and Walmart-Amazon (+3.1).
  • The method adds little computational overhead: training uses less memory and time per iteration than Hyp-ProtoNet, and inference adds about 4% memory, so the richer distance is practical.
  • Because both terms are differentiable, the distance can be used as a training signal for hyperbolic models that output sets, not just a fixed evaluation metric.

Reading between the lines

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

  • The tree-isomorphism theory in the paper assumes the two sets have equal size and a fixed ordering (the sequence indices are matched one-to-one), and that the graph is a tree; the implementation instead builds a complete weighted graph from all pairwise distances and applies the matrix sequence to it. If the four-word completeness assumption fails, the topological term should be read as a heuristi
  • A direct test of the approximation is to vary the number of Thue-Morse terms: if the topological term truly approximates an infinite polynomial basis, classification accuracy should saturate or improve with more terms and degrade when fewer than four are used.
  • The learned $\lambda$ adapter is a natural candidate for a controlled study: comparing HS2SD with a fixed optimal $\lambda$ would quantify how much of the gain comes from adaptivity as opposed to the topological term itself.
  • The paper's own stated limitation is the single-curvature assumption; an adaptive-curvature variant of HS2SD would be a direct next step that could be tested on the same few-shot and entity-matching benchmarks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes HS2SD, a hyperbolic set-to-set distance that combines a geodesic distance between Einstein midpoints (global structure) with a topological distance (local structure) computed from Thue-Morse words in the adjacency and degree matrices of a graph built from pairwise hyperbolic distances. The authors invoke McKay's spectral characterization of tree isomorphism (Theorem 4.1) and claim that the first four Thue-Morse words form a basis for representing tree topology (Corollary 4.5), leading to the topological distance of Eq. (8). The method is evaluated on unsupervised entity matching, standard image classification, and few-shot image classification, reporting consistent gains over point-to-point hyperbolic baselines.

Significance. If the theoretical claims were correct, HS2SD would be a useful contribution: a set-to-set distance in hyperbolic space with a principled decomposition into global and local terms, supported by a substantial appendix containing a proof of McKay's theorem and an analysis of Thue-Morse overlap-freeness. The empirical results are often strong (e.g., Table 4 shows 64.37% vs. 59.47% on miniImagenet 1-shot compared with Hyp-ProtoNet), and the paper includes ablations for each component. However, the central load-bearing claim—that a few Thue-Morse words span the matrix algebra of a tree—is not established and is contradicted by the paper's own Corollary D.26. Because the topological distance is the novel component of HS2SD, the reported gains cannot be attributed to the claimed mechanism as stated.

major comments (3)
  1. [§4.2, Corollary 4.5; Appendix E.3] Corollary 4.5 states that the first four Thue-Morse words form a basis for representing tree topology, but no proof is supplied and the claim is false as stated. For an asymmetric tree, Corollary D.26 correctly proves that the algebra generated by A and D is the full matrix algebra M_n(R), of dimension n^2; four fixed matrices (or five, if t4 is included as in Eq. (8)) cannot span this algebra for any n > 2. Appendix E.3 proves only that the Thue-Morse word is overlap-free and that any overlap-free binary morphism is a power of the Thue-Morse morphism (Theorem E.11); these are properties about redundancy, not about linear spanning. Proposition E.8 is the condition that would actually need to be verified, but the appendix never checks it for the words t0,...,t4. The theoretical justification of the topological distance therefore collapses.
  2. [§4.2.1; Eq. (8); Theorem 4.1] The graph used in the implementation is not a tree. Section 4.2.1 constructs the adjacency matrix A from all pairwise hyperbolic distances between feature points, producing a complete weighted graph. Theorem 4.1 and the McKay theorem (Theorem D.2/E.1) apply to unweighted trees, so the theoretical setting does not match the computation. In addition, Eq. (8) subtracts matrices t_i^x and t_i^y, which requires the two sets to have equal cardinality and a fixed ordering; in entity matching, the two records can have different numbers of attributes, and the Frobenius norm is then undefined. The paper does not explain how matrices of different sizes are aligned or padded, nor how the ordering of points within a set is determined.
  3. [§4.2, Eq. (8); Appendix E.2, Theorem E.6] Even if the words did span the algebra, the topological distance of Eq. (8) is an average of Frobenius norms of differences of individual words, whereas McKay's characterization is about cospectrality of every polynomial p(A,D), equivalently trace equality for every binary word. The proposed finite average of Frobenius norms is not equivalent to that spectral condition. Theorem E.6 establishes an isometry statement for the canonical homomorphism on the whole algebra, not for a finite set of words evaluated on two specific matrices; no error bound is proved for truncating the infinite family to five words. Thus the connection between Eq. (8) and Theorem 4.1 is not made, and the theoretical claim that d_t captures tree topology is unsupported.
minor comments (5)
  1. [§4.4, Eq. (10)] Eq. (10) applies a softmax inside the λ adapter, which outputs a vector, but λ in Eq. (6) is used as a scalar in [0,1]; the paper does not specify how the scalar is extracted from the softmax output.
  2. [§4.2, Corollary 4.5 vs. Eq. (8)] Corollary 4.5 says the first four terms t0,...,t3 are used, but Eq. (8) explicitly sums i=0..4, i.e., five terms. This inconsistency should be resolved.
  3. [§5.5, Table 6] The efficiency comparison in Table 6 uses different training configurations: Hyp-ProtoNet with 20-way training and HS2SD with 10-way training. Since the number of ways changes task difficulty and computational cost, the claimed memory and speed advantages are not an apples-to-apples comparison.
  4. [§5.4.2, Table 5] The ablation text says removing d_t costs 1.16% on tieredImagenet 1-shot (70.06% to 71.22%), but Table 5 lists the relevant value as 70.87% for CIFAR-FS and 70.87% for tieredImagenet; the numbers and dataset names do not match.
  5. [Abstract] The abstract contains a typo: 'The proposed the hyperbolic set-to-set distance measure' should read 'The proposed hyperbolic set-to-set distance measure'.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: HS2SD components are independently defined and benchmarked; the Thue-Morse basis gap is a soundness issue, not a circular reduction.

full rationale

The central HS2SD distance (Eq. 6) is evaluated against external benchmarks, and its components are not fitted to the target result: dg is the standard Einstein-midpoint/Mobius geodesic distance (Eq. 9), dt is a fixed average Frobenius norm over the first five Thue-Morse matrix words (Eq. 8), and the balance lambda is produced by a trained adapter (Eq. 10) with curvature c tuned on validation data. The claimed topological justification rests on McKay's theorem [47] and on Allouche and Shallit [1] via Corollary 4.5; these are external citations, not self-citations, and the Appendix re-derives McKay's theorem from standard algebra results rather than from the paper's own conclusions. The self-citations to [12], [18], and [46] appear only in related work and are not load-bearing. The serious weaknesses identified by the skeptic -- that four or five Thue-Morse words cannot span the n^2-dimensional algebra generated by A and D for asymmetric trees, that the implementation constructs complete weighted distance graphs rather than unweighted trees, and that Eq. (8) is not invariant under independent permutations of unordered set elements -- are conceptual and soundness gaps, not cases where a prediction reduces by construction to its input. No definitional circularity, fitted-input-called-prediction, or load-bearing self-citation chain is present. The score reflects only the presence of minor, non-load-bearing self-citations in the related-work section.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on McKay's tree-isomorphism theorem and on an unproved assertion that the first four Thue-Morse words span the matrix algebra. The implementation also assumes sets are trees of equal, ordered cardinality, which is not true for the entity-matching data. No new physical entities are introduced.

free parameters (3)
  • curvature c = 0.005 to 0.2; best 0.020 on AG (Fig 3)
    The hyperbolic distance formulas in Eqs. (5), (8), and (9) depend on c, and the paper sweeps c and selects the best on the AG dataset and validation sets.
  • balance lambda = 0.3 on AG; else produced by trained adapter G
    The relative weight of geodesic and topological components is either tuned as a scalar (Fig. 3) or learned by the lambda adapter, so it is fitted to data.
  • Thue-Morse depth n = 4
    The number of Thue-Morse terms is fixed at 4 by hand via Corollary 4.5, with no ablation over n and no proof that four terms span the algebra.
assumptions (5)
  • standard math McKay's tree-isomorphism theorem: two trees are isomorphic iff every polynomial p(A,D) is cospectral
    Restated as Theorem 4.1 and used as the theoretical foundation for the topological distance; cited from [47].
  • standard math Thue-Morse words are overlap-free and the overlap-free automatic morphism is essentially unique
    Used in Appendix E.3 to privilege Thue-Morse words; from [1].
  • domain assumption A hyperbolic set can be modeled as a tree whose topology is captured by adjacency and degree matrices
    Section 4.2 states sets are modeled as trees, but the implementation's adjacency matrix is built from all pairwise distances, yielding a complete weighted graph, not a tree.
  • ad hoc to paper The first four Thue-Morse words form a basis of the matrix algebra generated by A and D
    Corollary 4.5 asserts this without proof; Appendix E shows only overlap-freeness, and for an asymmetric tree the algebra has dimension n^2 (Corollary D.26).
  • ad hoc to paper The two sets have equal cardinality and a fixed ordering so that matrix subtraction in Eq. (8) is defined
    Eq. (8) subtracts matrices from the two sets; entity matching records have varying attribute counts, and no padding or alignment is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Set-to-Set Distance Measure in Hyperbolic Space." pith.science (2026). https://pith.science/paper/5QZMNW4X

@misc{pith2026250618529,
  author       = {Pith},
  title        = {Pith review of: A Set-to-Set Distance Measure in Hyperbolic Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5QZMNW4X}},
  note         = {Machine review of arXiv:2506.18529}
}
read the original abstract

We propose a hyperbolic set-to-set distance measure for computing dissimilarity between sets in hyperbolic space. While point-to-point distances in hyperbolic space effectively capture hierarchical relationships between data points, many real-world applications require comparing sets of hyperbolic data points, where the local structure and the global structure of the sets carry crucial semantic information. The proposed the \underline{h}yperbolic \underline{s}et-\underline{to}-\underline{s}et \underline{d}istance measure (HS2SD) integrates both global and local structural information: global structure through geodesic distances between Einstein midpoints of hyperbolic sets, and local structure through topological characteristics of the two sets. To efficiently compute topological differences, we prove that using a finite Thue-Morse sequence of degree and adjacency matrices can serve as a robust approximation to capture the topological structure of a set. In this case, by considering the topological differences, HS2SD provides a more nuanced understanding of the relationships between two hyperbolic sets. Empirical evaluation on entity matching, standard image classification, and few-shot image classification demonstrates that our distance measure outperforms existing methods by effectively modeling the hierarchical and complex relationships inherent in hyperbolic sets.

Figures

Figures reproduced from arXiv: 2506.18529 by the authors.

Figure 1
Figure 1. Examples of HS2SD on the entity matching and few-shot learning tasks. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of Hyperbolic Set-to-Set Distance Measure. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation study on curvature c and lambda λ on AG dataset for EM task. 5.4.4 Ablations on the curvature c and the λ of HS2SD Impact of Curvature [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 65 canonical work pages

  1. [1]

    Cambridge university press, 2003

    Jean-Paul Allouche and Jeffrey Shallit.Automatic sequences: theory, applications, generalizations. Cambridge university press, 2003

    work page 2003

  2. [2]

    Gpr-net: Geometric prototypical network for point cloud few-shot learning

    Tejas Anvekar and Dena Bazazian. Gpr-net: Geometric prototypical network for point cloud few-shot learning. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4179–4188, 2023

    work page 2023

  3. [3]

    Hyperbolic image segmentation

    Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne van Noord, and Pascal Mettes. Hyperbolic image segmentation. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4453–4462, June 2022

    work page 2022

  4. [4]

    Hyperbolic image segmentation

    Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4453–4462, 2022

    work page 2022

  5. [5]

    Multi-relational poincaré graph embeddings

    Ivana Balazevic, Carl Allen, and Timothy Hospedales. Multi-relational poincaré graph embeddings. 32, 2019

    work page 2019

  6. [6]

    Fully hyperbolic convolutional neural networks for computer vision

    Ahmad Bdeir, Kristian Schwethelm, and Niels Landwehr. Fully hyperbolic convolutional neural networks for computer vision. InInternational Conference on Learning Representations (ICLR), 2024

    work page 2024

  7. [7]

    Meta-learning with differentiable closed-form solvers.arXiv preprint arXiv:1805.08136, 2018

    Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers.arXiv preprint arXiv:1805.08136, 2018

    arXiv 2018

  8. [8]

    Number 67

    Norman Biggs.Algebraic graph theory. Number 67. Cambridge university press, 1993

    work page 1993

Show all 76 references
  1. [9]

    Creating embeddings of heteroge- neous relational datasets for data integration tasks

    Riccardo Cappuzzo, Paolo Papotti, and Saravanan Thirumuruganathan. Creating embeddings of heteroge- neous relational datasets for data integration tasks. InProceedings of the 2020 ACM SIGMOD international conference on management of data, pages 1335–1349, 2020

    2020

  2. [10]

    Hyperbolic uncertainty aware semantic segmenta- tion.arXiv preprint arXiv:2203.08881, 2022

    Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmenta- tion.arXiv preprint arXiv:2203.08881, 2022

    2022 arXiv

  3. [11]

    Hyperbolic uncertainty aware semantic segmenta- tion.IEEE Transactions on Intelligent Transportation Systems (T-ITS), 2023

    Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmenta- tion.IEEE Transactions on Intelligent Transportation Systems (T-ITS), 2023

    2023

  4. [12]

    A hyperbolic-to-hyperbolic graph convolutional network

    Jindou Dai, Yuwei Wu, Zhi Gao, and Yunde Jia. A hyperbolic-to-hyperbolic graph convolutional network. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 154–163, 2021

    2021

  5. [13]

    Hyperbolic vision transformers: Combining improvements in metric learning

    Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7409–7419, June 2022

    2022

  6. [14]

    Hyperbolic vision transformers: Combining improvements in metric learning

    Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7409–7419, 2022

    2022

  7. [15]

    American Mathematical Soc., 2011

    Pavel I Etingof, Oleg Golberg, Sebastian Hensel, Tiankai Liu, Alex Schwendner, Dmitry Vaintrob, and Elena Yudovina.Introduction to representation theory, volume 59. American Mathematical Soc., 2011. 10

    2011

  8. [16]

    Kernel methods in hyperbolic spaces

    Pengfei Fang, Mehrtash Harandi, and Lars Petersson. Kernel methods in hyperbolic spaces. InInternational Conference on Computer Vision (ICCV), pages 10665–10674, 2021

    2021

  9. [17]

    Hyperbolic neural networks

    Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. 31, 2018

    2018

  10. [18]

    Curvature generation in curved spaces for few-shot learning

    Zhi Gao, Yuwei Wu, Yunde Jia, and Mehrtash Harandi. Curvature generation in curved spaces for few-shot learning. InInternational Conference on Computer Vision (ICCV), pages 8691–8700, 2021

    2021

  11. [19]

    Hyperbolic busemann learning with ideal prototypes

    Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes. 34:103–115, 2021

    2021

  12. [20]

    Some computational results on the spectra of graphs

    Chris Godsil and Brendan Mckay. Some computational results on the spectra of graphs. InCombinatorial Mathematics IV: Proceedings of the Fourth Australian Conference Held at the University of Adelaide August 27–29, 1975, pages 73–92. Springer, 1976

    1975

  13. [21]

    Springer, 1987

    Mikhael Gromov.Hyperbolic groups. Springer, 1987

    1987

  14. [22]

    Hyperbolic attention networks

    Caglar Gulcehre, Misha Denil, Mateusz Malinowski, Ali Razavi, Razvan Pascanu, Karl Moritz Hermann, Peter Battaglia, Victor Bapst, David Raposo, Adam Santoro, and Nando de Freitas. Hyperbolic attention networks. InInternational Conference on Learning Representations (ICLR), 2019

    2019

  15. [23]

    Clipped hyperbolic classifiers are super- hyperbolic classifiers

    Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super- hyperbolic classifiers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11–20, 2022

    2022

  16. [24]

    Camper: An effective framework for privacy-aware deep entity resolution

    Yuxiang Guo, Lu Chen, Zhengjie Zhou, Baihua Zheng, Ziquan Fang, Zhikun Zhang, Yuren Mao, and Yunjun Gao. Camper: An effective framework for privacy-aware deep entity resolution. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 626–637, 2023

    2023

  17. [25]

    Springer Science & Business Media, 2013

    Gena Hahn and Gert Sabidussi.Graph symmetry: algebraic methods and applications, volume 497. Springer Science & Business Media, 2013

    2013

  18. [26]

    Hyperbolic prototypical network for few shot remote sensing scene classification.Pattern Recognition (PR), 177:151–156, 2024

    Manal Hamzaoui, Laetitia Chapel, Minh-Tan Pham, and Sébastien Lefèvre. Hyperbolic prototypical network for few shot remote sensing scene classification.Pattern Recognition (PR), 177:151–156, 2024

    2024

  19. [27]

    Capturing implicit hierarchical structure in 3d biomedical images with self-supervised hyperbolic representations

    Joy Hsu, Jeffrey Gu, Gong Her Wu, Wah Chiu, and Serena Yeung. Capturing implicit hierarchical structure in 3d biomedical images with self-supervised hyperbolic representations. 2021

    2021

  20. [28]

    A localized geometric method to match knowledge in low- dimensional hyperbolic space

    Bo Hui, Tian Xia, and Wei-Shinn Ku. A localized geometric method to match knowledge in low- dimensional hyperbolic space. InConference on Empirical Methods in Natural Language Processing (EMNLP), pages 2822–2832, 2022

    2022

  21. [29]

    On line assemblies.Journal für die reine und angewandte Mathematik (in French), 70(2):185–190, 1869

    Camille Jordan. On line assemblies.Journal für die reine und angewandte Mathematik (in French), 70(2):185–190, 1869

  22. [30]

    Hyperbolic image embeddings

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6418–6428, 2020

    2020

  23. [31]

    On the solution of the equations obtained from the investigation of the linear distribution of galvanic currents.IRE transactions on circuit theory, 5(1):4–7, 1958

    Gustav Kirchhoff. On the solution of the equations obtained from the investigation of the linear distribution of galvanic currents.IRE transactions on circuit theory, 5(1):4–7, 1958

    1958

  24. [32]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  25. [33]

    Courier Corporation, 2012

    Lauwerens Kuipers and Harald Niederreiter.Uniform distribution of sequences. Courier Corporation, 2012

    2012

  26. [34]

    A basis theorem for matrix algebras.Linear and Multilinear Algebra, 8(3):183–187, 1980

    Thomas J Laffey. A basis theorem for matrix algebras.Linear and Multilinear Algebra, 8(3):183–187, 1980

    1980

  27. [35]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  28. [36]

    MNIST handwritten digit database

    Yann LeCun and Corinna Cortes. MNIST handwritten digit database. 2010

    2010

  29. [37]

    Meta-learning with differen- tiable convex optimization

    Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differen- tiable convex optimization. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10657–10665, 2019. 11

    2019

  30. [38]

    Neighborhood-adaptive structure augmented metric learning.AAAI Conference on Artificial Intelligence (AAAI), 36(2):1367–1375, Jun

    Pandeng Li, Yan Li, Hongtao Xie, and Lei Zhang. Neighborhood-adaptive structure augmented metric learning.AAAI Conference on Artificial Intelligence (AAAI), 36(2):1367–1375, Jun. 2022

    2022

  31. [39]

    Auto-fuzzyjoin: Auto-program fuzzy similarity joins without labeled examples

    Peng Li, Xiang Cheng, Xu Chu, Yeye He, and Surajit Chaudhuri. Auto-fuzzyjoin: Auto-program fuzzy similarity joins without labeled examples. InProceedings of the 2021 international conference on management of data, pages 1064–1076, 2021

    2021

  32. [40]

    Mhcn: A hyperbolic neural network model for multi-view hierarchical clustering

    Fangfei Lin, Bing Bai, Yiwen Guo, Hao Chen, Yazhou Ren, and Zenglin Xu. Mhcn: A hyperbolic neural network model for multi-view hierarchical clustering. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16525–16535, October 2023

    2023

  33. [41]

    Hyperbolic chamfer distance for point cloud completion

    Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, and Ziming Zhang. Hyperbolic chamfer distance for point cloud completion. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14595–14606, 2023

    2023

  34. [42]

    Hyperbolic visual embedding learning for zero-shot recognition

    Shaoteng Liu, Jingjing Chen, Liangming Pan, Chong-Wah Ngo, Tat-Seng Chua, and Yu-Gang Jiang. Hyperbolic visual embedding learning for zero-shot recognition. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9273–9281, 2020

    2020

  35. [43]

    Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692, 364, 2019

    Yinhan Liu. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692, 364, 2019

    1907 arXiv

  36. [44]

    Searching for actions on the hyperbole

    Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1141–1150, 2020

    2020

  37. [45]

    Cross-modal scalable hyperbolic hierarchical clustering

    Teng Long and Nanne van Noord. Cross-modal scalable hyperbolic hierarchical clustering. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16655–16664, October 2023

    2023

  38. [46]

    Adaptive poincaré point to set distance for few-shot classification

    Rongkai Ma, Pengfei Fang, Tom Drummond, and Mehrtash Harandi. Adaptive poincaré point to set distance for few-shot classification. InAAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 1926–1934, 2022

    1926

  39. [47]

    On the spectral characterisation of trees.Ars Combin, 3:219–232, 1977

    Brendan D McKay. On the spectral characterisation of trees.Ars Combin, 3:219–232, 1977

    1977

  40. [48]

    Recurrent geodesics on a surface of negative curvature.Transactions of the American Mathematical Society, 22(1):84–100, 1921

    Harold Marston Morse. Recurrent geodesics on a surface of negative curvature.Transactions of the American Mathematical Society, 22(1):84–100, 1921

    1921

  41. [49]

    Deep learning for entity matching: A design space exploration

    Sidharth Mudgal, Han Li, Theodoros Rekatsinas, AnHai Doan, Youngchoon Park, Ganesh Krishnan, Rohit Deep, Esteban Arcaute, and Vijay Raghavendra. Deep learning for entity matching: A design space exploration. InProceedings of the 2018 international conference on management of d...

    2018

  42. [50]

    Tadam: Task dependent adaptive metric for improved few-shot learning.Advances in neural information processing systems, 31, 2018

    Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning.Advances in neural information processing systems, 31, 2018

    2018

  43. [51]

    Spectral clus- tering with anchor graph based on set-to-set distances for large-scale hyperspectral images.International Journal of Remote Sensing, 43(7):2438–2460, 2022

    Yao Qin, Sinong Quan, Chongyang Wei, Weiping Ni, Kun Li, Xiaohu Dong, and Yuanxin Ye. Spectral clus- tering with anchor graph based on set-to-set distances for large-scale hyperspectral images.International Journal of Remote Sensing, 43(7):2438–2460, 2022

    2022

  44. [52]

    Meta-learning for semi-supervised few-shot classification.arXiv preprint arXiv:1803.00676, 2018

    Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification.arXiv preprint arXiv:1803.00676, 2018

    2018 arXiv

  45. [53]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge.International Journal of Computer Vision (IJCV...

    2015

  46. [54]

    Representation tradeoffs for hyperbolic embeddings

    Frederic Sala, Chris De Sa, Albert Gu, and Christopher Re. Representation tradeoffs for hyperbolic embeddings. InInternational Conference on Machine Learning (ICML), volume 80 ofProceedings of Machine Learning Research, pages 4460–4469. PMLR, 10–15 Jul 2018

    2018

  47. [55]

    Privacy-preserving record linkage using bloom filters

    Rainer Schnell, Tobias Bachteler, and Jörg Reiher. Privacy-preserving record linkage using bloom filters. BMC medical informatics and decision making, 9:1–11, 2009

    2009

  48. [56]

    University of Michigan, 1973

    Allen John Carl Schwenk.The spectrum of a graph. University of Michigan, 1973. 12

    1973

  49. [57]

    Hyperbolic neural networks++.arXiv preprint arXiv:2006.08210, 2020

    Ryohei Shimizu, Yusuke Mukuta, and Tatsuya Harada. Hyperbolic neural networks++.arXiv preprint arXiv:2006.08210, 2020

    2006 arXiv

  50. [58]

    Prototypical networks for few-shot learning.Advances in neural information processing systems, 30, 2017

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.Advances in neural information processing systems, 30, 2017

    2017

  51. [59]

    Domain adaptation for deep entity resolution

    Jianhong Tu, Ju Fan, Nan Tang, Peng Wang, Chengliang Chai, Guoliang Li, Ruixue Fan, and Xiaoyong Du. Domain adaptation for deep entity resolution. InProceedings of the 2022 International Conference on Management of Data, pages 443–457, 2022

    2022

  52. [60]

    Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry.Computers & Mathematics with Applications, 41(1-2):135–147, 2001

    Abraham A Ungar. Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry.Computers & Mathematics with Applications, 41(1-2):135–147, 2001

    2001

  53. [61]

    Which graphs are determined by their spectrum?Linear Algebra and its applications, 373:241–272, 2003

    Edwin R Van Dam and Willem H Haemers. Which graphs are determined by their spectrum?Linear Algebra and its applications, 373:241–272, 2003

    2003

  54. [62]

    Poincaré resnet

    Max van Spengler, Erwin Berkhout, and Pascal Mettes. Poincaré resnet. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5419–5428, 2023

    2023

  55. [63]

    Matching networks for one shot learning

    Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. 29, 2016

    2016

  56. [64]

    Spd manifold deep metric learning for image set classification.IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2024

    Rui Wang, Xiao-Jun Wu, Ziheng Chen, Cong Hu, and Josef Kittler. Spd manifold deep metric learning for image set classification.IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2024

    2024

  57. [65]

    Negative sample matters: A renaissance of metric learning for temporal grounding

    Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. InAAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 2613–2623, 2022

    2022

  58. [66]

    Zeroer: En- tity resolution using zero labeled examples

    Renzhi Wu, Sanya Chaba, Saurabh Sawlani, Xu Chu, and Saravanan Thirumuruganathan. Zeroer: En- tity resolution using zero labeled examples. InProceedings of the 2020 ACM SIGMOD International Conference on Management of Data, pages 1149–1164, 2020

    2020

  59. [67]

    Deep features for person re-identification on metric learning.Pattern Recognition (PR), 110:107424, 2021

    Wanyin Wu, Dapeng Tao, Hao Li, Zhao Yang, and Jun Cheng. Deep features for person re-identification on metric learning.Pattern Recognition (PR), 110:107424, 2021

    2021

  60. [68]

    Unsupervised hyperbolic metric learning

    Jiexi Yan, Lei Luo, Cheng Deng, and Heng Huang. Unsupervised hyperbolic metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12465–12474, 2021

    2021

  61. [69]

    Hyp-uml: Hyperbolic image retrieval with uncertainty-aware metric learning.arXiv preprint arXiv:2310.08390, 2023

    Shiyang Yan, Zongxuan Liu, and Lin Xu. Hyp-uml: Hyperbolic image retrieval with uncertainty-aware metric learning.arXiv preprint arXiv:2310.08390, 2023

    2023 arXiv

  62. [70]

    Few-shot learning via embedding adaptation with set-to-set functions

    Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. InThe IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8808–8817, 2020

    2020

  63. [71]

    Few-shot learning via embedding adaptation with set-to-set functions

    Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8808–8817, 2020

    2020

  64. [72]

    Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    2016 arXiv

  65. [73]

    Hyperbolic knowl- edge transfer with class hierarchy for few-shot learning

    Baoquan Zhang, Hao Jiang, Shanshan Feng, Xutao Li, Yunming Ye, and Rui Ye. Hyperbolic knowl- edge transfer with class hierarchy for few-shot learning. InInternational Joint Conference on Artificial Intelligence (IJCAI), pages 3723–3729, 7 2022. Main Track

    2022

  66. [74]

    Hyperbolic knowledge transfer with class hierarchy for few-shot learning

    Baoquan Zhang, Hao Jiang, Shanshan Feng, Xutao Li, Yunming Ye, and Rui Ye. Hyperbolic knowledge transfer with class hierarchy for few-shot learning. InIJCAI, pages 3723–3729, 2022

    2022

  67. [75]

    determined by

    Pengfei Zhu, Mingqi Gu, Wenbin Li, Changqing Zhang, and Qinghua Hu. Progressive point to set metric learning for semi-supervised few-shot classification. InIEEE International Conference on Image Processing (ICIP), pages 196–200, 2020. 13 A Broader Impact The ability to leverag...

    2020

  68. [76]

    µ is called the Thue-Morse morphism

    For example, let µ(D) =DA, µ(A) =AD , then the Thue-Morse words are the iterations µi(D). µ is called the Thue-Morse morphism. We call h overlap free, if h(w)is overlap free for any overlap free wordw. Then Theorem E.11 For any homomorphism h on binary words in A and D, if h(D...

Pith tools

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