Pith. sign in

REVIEW 3 major objections 4 minor 68 references

Unsupervised Ground Metric Learning

T0 review · 3 major / 4 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 Stochastic coordinate updates converge linearly for unsupervised metric learning.

desk verdict Theorem 6 is a genuine theoretical advance and the proof checks out; but the experiments run far outside its step-size regime, and the paper admits this. read the letter →

arxiv 2507.13094 v1 pith:ZFWA3WOG submitted 2025-07-17 math.OC cs.LGcs.NAmath.NA

Janis Auffenberg , Jonas Bresch , Oleh Melnyk , Gabriele Steidl This is my paper
classification math.OCcs.LGcs.NAmath.NA MSC 49Q2265K10
keywords unsupervisedmetriclearningoptimaltransportSinkhorndivergencestochasticfixedpointiterationlinearconvergenceMahalanobisdistancegraphLaplacianPerron-Frobenius
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 aims to put unsupervised metric learning, where feature distances are learned without labels, on firmer theoretical ground. It studies two coupled fixed-point problems in which a distance matrix and a cost matrix define each other through optimal transport or related distances, and it proposes deterministic and stochastic iterations to solve them. For the stochastic Random Function Iteration algorithm, it proves linear convergence, both in expectation and almost surely, to a unique fixed point under Lipschitz conditions alone, without the paracontractivity assumption that earlier convergence results required. It also replaces the optimal transport distance by Sinkhorn divergences, Mahalanobis-like kernel distances, and graph Laplacian distances, showing that the last option reduces to a linear eigenvalue problem. If the proof carries, a cheap single-entry stochastic update with constant step sizes can replace full matrix evaluations in these metric learning tasks.

What carries the argument

The load-bearing object is Algorithm 3, the Random Function Iteration scheme, together with the coupled coordinate updates $R^{(i,j)}$ and $S^{(k,\ell)}$ that move one entry of $B$ toward $\gamma_F F(A)$ and one entry of $A$ toward $\gamma_G G(B)$. The proof's engine is the averaging step: taking expectation over the random index turns the coordinate-update operator $T^\xi$ into a contraction in the $\ell^2$ metric, yielding the factor $L < 1$. Equally important is the decision to measure errors in the $\ell^2$ norm rather than $\ell^\infty$; in $\ell^\infty$ the single-entry updates can stall at points of equal error, which is exactly why previous paracontractivity results do not apply.

What would settle it

Check whether Theorem 6's contraction bound holds when the Lipschitz constants are exactly at the boundary values $L_F = \sqrt{\alpha}/(\sqrt{2m}\gamma_F)$ and $L_G = 1/(n\gamma_G)$: for fixed small $m,n$ and each coordinate choice, compute the one-step expected squared error in the $\ell^2$ norm. A single counterexample would refute the theorem; alternatively, running Algorithm 3 with step sizes one order of magnitude above the stated bounds on the PCA-reduced single-cell dataset would show whether the dimension-dependent smallness condition is actually necessary for convergence.

Watch

Extended reading notes

Core claim

The central theoretical discovery is that the stochastic Random Function Iteration algorithm, which updates one entry of the matrix pair per step, converges linearly to the unique solution of the coupled system $B = \gamma_F F(A)$, $A = \gamma_G G(B)$, even though the operators are nonexpansive but not paracontractive. Theorem 6 shows that if $F$ and $G$ are Lipschitz with constants bounded by $L_F \leq \sqrt{\alpha}/(\sqrt{2m}\gamma_F)$ and $L_G \leq 1/(n\gamma_G)$, then the expected squared distance to the fixed point contracts by a factor $L < 1$ each iteration, and the same factor implies almost sure convergence. The proof works by averaging over the random coordinate choices: each single-coordinate update is only nonexpansive, but in expectation it behaves like a paracontraction when measured in the $\ell^2$ norm. The paper further proves matching convergence for the normalized fixed-point problem, establishes existence of fixed points for Sinkhorn and Wasserstein maps, and shows that the graph-Laplacian variant reduces to a Perron-Frobenius eigenvalue problem.

Load-bearing premise

The linear-convergence guarantee holds only when the step sizes are small enough relative to the Lipschitz constants and the data dimensions ($\gamma_F \lesssim 1/\sqrt{m}$, $\gamma_G \lesssim 1/n$), while the experiments that achieve strong clustering use step sizes far outside this range, so the theorem does not cover the practically successful regime.

Editorial extensions

If this is right

  • For the Wasserstein and Sinkhorn settings, the stochastic iteration converges almost surely with constant step sizes $\gamma_F \leq \sqrt{\alpha}/\sqrt{2m}$ and $\gamma_G \leq 1/n$, so no vanishing step-size schedule is needed.
  • The iterate sequence is a supermartingale whose expected squared error decreases by a fixed factor each step, giving a worst-case linear rate on the order of $1 - 1/n^4$ or $1 - \alpha/(2m^2)$.
  • The same convergence framework covers Mahalanobis-like distances built from radial positive-definite kernels, provided their Lipschitz constants satisfy the same bounds.
  • For the graph-Laplacian model, both maps are linear, so the fixed point becomes the largest eigenvector of a matrix with strictly positive entries; Perron-Frobenius guarantees a unique positive eigenvector and hence a metric matrix.
  • The stochastic iteration avoids the normalization step used in earlier stochastic power iterations, which the paper identifies as unnecessary for convergence.
  • All proof fields end with complete sentences, as required.

Reading between the lines

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

  • The dimension-dependent step-size bounds in Theorem 6 are likely far from sharp; the experiments show convergence and strong clustering for step sizes far outside the stated range, suggesting the Lipschitz-based smallness conditions are sufficient but not necessary.
  • The paper's observation that the normalized algorithm does not converge yet gives the best clustering suggests that early stopping acts as regularization; this could be tested by comparing cluster quality at the fixed point versus at the residual peak.
  • The graph-Laplacian reduction to a linear eigenvalue problem opens a direct route to spectral clustering with learned metrics, because the learned matrix is guaranteed to be a metric matrix.
  • The adaptive step-size experiments indicate that the fixed point of the scaled system depends continuously on $(\gamma_F, \gamma_G)$, so adaptively chosen step sizes could replace the normalization in Algorithm 1; this is a testable hypothesis but not a claim of the paper.
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 / 4 minor

Summary. The paper studies unsupervised ground metric learning through two coupled fixed-point problems: the normalized problem (2) and the scaled problem (3) for maps F and G between matrix spaces. It proposes a deterministic iteration for each problem and a stochastic Random Function Iteration algorithm (Algorithm 3) for problem (3). The central theoretical contribution is Theorem 6, which proves linear convergence in expectation and almost sure convergence of Algorithm 3 to the unique fixed point under Lipschitz conditions on F and G, without assuming paracontractivity of the individual update operators. The paper then instantiates F and G for optimal transport distances, Sinkhorn divergences, Mahalanobis-like kernel distances, and graph Laplacian distances, with existence and convergence results for each model. Numerical experiments on synthetic translated histograms and on a scRNA-seq dataset compare the proposed methods with Euclidean distance and with the Wasserstein singular vector method of Huizing et al.

Significance. If the main results are correct, the paper makes a useful theoretical contribution: it extends convergence guarantees for stochastic random function iterations from paracontractive to merely nonexpansive-on-average operators, and it broadens unsupervised metric learning beyond the optimal-transport setting to Mahalanobis-type and graph-Laplacian models. The proof of Theorem 6 is a coherent supermartingale argument that I checked in detail through equations (10) and (11) and found internally consistent. The paper also provides explicit existence proofs via Brouwer's theorem and Perron-Frobenius theory, and releases code for the experiments. However, two load-bearing statements are not correct as written: the contraction condition in Theorem 1 is not sufficient for the claimed conclusion, and the graph Laplacian section claims positive definiteness for objects that are singular by construction. In addition, the numerical experiments deliberately operate in a parameter regime far outside the assumptions of Theorem 6, and this gap is not flagged clearly enough.

major comments (3)
  1. [Section 2.1, Theorem 1] The contraction condition in Theorem 1 is not sufficient.
  2. [Section 5, Lemma 21 and Theorem 22] The graph Laplacian section uses an inconsistent notion of positive definiteness.
  3. [Section 6 and Theorem 6] The experimental regime is not covered by Theorem 6.
minor comments (4)
  1. [Section 3, Theorem 10.3] Theorem 10.3 states that Algorithm 3 converges to the unique fixed point of problem (2), but Algorithm 3 is designed for problem (3); the reference should be to (3).
  2. [Section 6, Figure 3 caption] The caption of Figure 3 refers to 'methods from Fig. 3', which is self-referential and should probably refer to Figure 1 or Figure 2.
  3. [Acknowledgments] The acknowledgment thanks 'Geert-Jan Huitzing', but the cited author of reference [33] is spelled 'Huizing'; the spelling should be made consistent.
  4. [Section 3, Theorem 13 proof] In the proof of Theorem 13, the bound on \(W^\varepsilon_{0_{n,n}}\) contains \(\varepsilon \|A\|_\infty\), which should be \(\varepsilon \|0_{n,n}\|_\infty = 0\), since the cost matrix is the zero matrix.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the convergence proofs are self-contained; the experimental regime gap is explicitly acknowledged and separated from the proven claims.

full rationale

The derivation chain is self-contained at every load-bearing step. The fixed-point problems (2)-(3) are definitions, and Algorithms 1-3 are direct iteration schemes. The main convergence results (Theorems 1, 2, 6) are proved from Banach's fixed point theorem and elementary expectation estimates: Theorem 6 computes a contraction factor L from the step-size bounds gamma_F <= sqrt(alpha)/(sqrt(2m) L_F) and gamma_G <= 1/(n L_G), and the bound L < 1 follows algebraically from these inequalities (eqs. (10)-(11) and the final estimate on L in the proof). No quantity in L is fitted to numerical output, and no convergence claim is asserted from empirical success. The model-specific Lipschitz constants used in Theorems 10, 13, 19, 22 are derived from the definitions of Wasserstein, Sinkhorn, Mahalanobis, and graph-Laplacian maps (Lemmas 9, 12, 16) rather than assumed. Existence of fixed points is established via Brouwer (Theorem 13) and Perron-Frobenius (Theorem 22), with explicit hypotheses (e.g., positivity of the matrix H in Theorem 22 is proven from assumption (22)). Citations to prior work are not load-bearing for the central claims: references [2]-[4] of co-author Bresch appear only as related-work pointers, and the appendix's use of Lemma 28 from [33] concerns an optional existence variant, not the main theorem chain. The paper also explicitly flags that its theorem conditions are not met by several reported experiments: Section 6.2 states that 'conditions in Theorem 10 do not apply' for SEVn, and the Conclusions states that 'Algorithm 1 converges in scenarios beyond those covered by theoretical results. It remains an open problem to justify its performance.' These admissions are the opposite of circularity: they separate the proven regime from the empirical regime. Therefore no 'prediction' reduces to an input by construction, and the circularity score is 0.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central theorems use standard fixed-point results (Banach, Brouwer, Perron-Frobenius) and a supermartingale convergence argument. The application sections impose domain assumptions (normalization of X, positive data entries for Sinkhorn, radial positive definiteness of kernels, and the genericity condition (22) for the graph Laplacian). No new entities are postulated; all quantities are derived from the data X and the chosen reference matrices.

free parameters (5)
  • Reference matrices R_n, R_m = τ R with τ=1e-2 or 1e-3
    Added to F and G in (14), (17), (18) to ensure positivity and boundedness of ||F(A)||_∞ (Lemmas 9, 12, 17). Values are chosen by hand in the experiments.
  • Step sizes γ_F, γ_G = 0.9, 1.0, or adaptive
    Control the fixed-point update in (3). In experiments, γ=0.9 is used for SEV even though the stochastic bound in Theorem 6 requires γ_F ≤ √α/√(2m) and γ_G ≤ 1/n, which are not satisfied.
  • Averaging parameter α = 0.9
    Appears in operator T in (4); set to 0.9 in experiments.
  • Sinkhorn regularization ε = 5e-2 (synthetic), 1e-1 (scRNA-seq)
    Entropic regularization parameter in the Sinkhorn divergence (16); chosen by hand.
  • Kernel widths σ_F, σ_G = 1 (synthetic), 10 and 1 (scRNA-seq)
    Widths of the Gaussian or Laplacian radial functions in Section 4; chosen by hand. Unreported grid search showed little impact.
assumptions (7)
  • standard math Banach fixed point theorem
    Used in Theorems 1, 2, and Lemma 4 to obtain existence and uniqueness of fixed points.
  • standard math Brouwer fixed point theorem
    Used in Theorem 13 to prove existence of a fixed point of ~T.
  • standard math Perron-Frobenius theorem
    Used in Theorem 22 to obtain a positive eigenvector for the linear graph Laplacian operator.
  • domain assumption Data normalization X 1_n = 1_m and X^T 1_m = 1_n
    Assumed in Section 3 for OT and Sinkhorn distances so that rows and columns are probability vectors.
  • domain assumption Radially positive definite kernel functions f,g (inequality (19))
    Used in Lemma 15 to guarantee that F and G map into the positive semidefinite cone.
  • ad hoc to paper Genericity assumption (22) for the graph Laplacian data
    Ensures the matrix H in Theorem 22 has positive entries so Perron-Frobenius applies.
  • domain assumption Sinkhorn data entries are strictly positive for the Lipschitz constant C to be finite
    Lemma 12 needs X_{i,k} > 0 wherever the log terms are evaluated; in experiments this is achieved by an exponential transform.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Ground Metric Learning." pith.science (2026). https://pith.science/paper/ZFWA3WOG

@misc{pith2026250713094,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Ground Metric Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZFWA3WOG}},
  note         = {Machine review of arXiv:2507.13094}
}
read the original abstract

Data classification without access to labeled samples remains a challenging problem. It usually depends on an appropriately chosen distance between features, a topic addressed in metric learning. Recently, Huizing, Cantini and Peyr\'e proposed to simultaneously learn optimal transport (OT) cost matrices between samples and features of the dataset. This leads to the task of finding positive eigenvectors of a certain nonlinear function that maps cost matrices to OT distances. Having this basic idea in mind, we consider both the algorithmic and the modeling part of unsupervised metric learning. First, we examine appropriate algorithms and their convergence. In particular, we propose to use the stochastic random function iteration algorithm and prove that it converges linearly for our setting, although our operators are not paracontractive as it was required for convergence so far. Second, we ask the natural question if the OT distance can be replaced by other distances. We show how Mahalanobis-like distances fit into our considerations. Further, we examine an approach via graph Laplacians. In contrast to the previous settings, we have just to deal with linear functions in the wanted matrices here, so that simple algorithms from linear algebra can be applied.

Figures

Figures reproduced from arXiv: 2507.13094 by the authors.

Figure 1
Figure 1. Comparison of the considered eigenvalue methods. Here, the resulting dis￾tances distpXi , Xj q are depicted for the dataset Xp3q . We distinguish two variants for each of the algorithms. The first is a non-normalized with mapping T as in (4) with α “ 0.9 and γF “ γG “ 0.75 for the WEV and SEV and the RBF-MSEV and GMEV with α “ 0.9 and γF “ γG “ 0.01. The second is a normalized iteration corresponding to the mapping … view at source ↗
Figure 2
Figure 2. Distance distpX p3q 20 , Xp3q i q for the normalized (left) and non-normalized (right) eigenvector computation schemes. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. ℓ8-residual and Hilbert norm progression with iterations for the metric learn￾ing methods from [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: t-SNE plots of the resulting cell clustering using learned metrics on the PCA-reduced scRNAseq dataset. Turning to OT-based distances, we again observe a dichotomy between the performance on the reduced and full datasets. On the reduced dataset, all Sinkhorn-based meth…
Figure 5
Figure 5. Figure 5: Evolution of ASW and the ℓ8-residual with iteration for optimal transport methods for PCA-reduced (top) and full (bottom) scRNAseq datasets. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of cell clusters with learned distances on the full scR￾NAseq dataset. cells from Mono and B cells. Time-wise, SEV with γ “ 0.9 and its stochastic version are the fastest. Note that we stopped sSEV after 54 iterations at the peak ASW shown in [PITH…
Figure 7
Figure 7. Figure 7: t-SNE visualization of the marker gene clusters using learned distances on the full scRNAseq dataset. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_7.png]
Figure 8
Figure 8. Figure 8: ASW development for SEVn and SEV for different choices of γF and γG and adaptive updates according to the ℓ8-residual, where τ “ 10´3 is constant. 1 3 5 7 9 11 13 15 0 0.2 0.4 0.6 ◦ SEVn × SEV τ = 10−1 τ = 10−3 τ = 10−5 [PITH_FULL_IMAGE:figures/full_fig_p034_8.png]
Figure 10
Figure 10. Figure 10: t-SNE visualization of the resulting clustering of cells (top) and marker genes (bottom) using learned distances on the full scRNAseq dataset. Left: SEVn with τ “ 10´5 , center: early stopped SEV with τ “ 10´5 , right: SEV with adaptive step size and τ “ 10´3 . Impact…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 46 canonical work pages

  1. [33]

    Unsupervised Ground Met- ric Learning Using Wasserstein Singular Vectors

    Geert-Jan Huizing, Laura Cantini, and Gabriel Peyr´ e. “Unsupervised Ground Met- ric Learning Using Wasserstein Singular Vectors”. In: Proceedings of the 39th International Conference on Machine Learning . Ed. by Kamalika Chaudhuri et al. Vol. 162. Proceedings of Machine Learning Research. PMLR, 17–23 Jul 2022, pp. 9429–9443. url: https://proceedings.mlr....

    work page 2022

  2. [1]

    Sur les op´ erations dans les ensembles abstraits et leur application aux ´ equations int´ egrales

    Stefan Banach. “Sur les op´ erations dans les ensembles abstraits et leur application aux ´ equations int´ egrales”. In:Fundamenta Mathematicae 3 (1922), pp. 133–181. doi: 10.4064/fm-3-1-133-181

    doi:10.4064/fm-3-1-133-181 1922

  3. [2]

    Max-Normalized Radon Cumulative Distribution Transform for Limited Data Classification

    Matthias Beckmann, Robert Beinert, and Jonas Bresch. “Max-Normalized Radon Cumulative Distribution Transform for Limited Data Classification”. In: Scale Space and Variational Methods in Computer Vision . Vol. 15667. Lecture Notes in Computer Science. Springer, 2025, pp. 241–254. doi: 10.1007/978- 3- 031- 92366-1_19

    doi:10.1007/978- 2025

  4. [3]

    Normalized Radon Cumulative Distribution Transforms for Invariance and Robustness in Optimal Transport Based Image Classification

    Matthias Beckmann, Robert Beinert, and Jonas Bresch. “Normalized Radon Cu- mulative Distribution Transforms for Invariance and Robustness in Optimal Trans- port Based Image Classification”. 2025. doi: 10.48550/arXiv.2506.08761

    work page Pith review arXiv doi:10.48550/arxiv.2506.08761 2025

  5. [4]

    Joint Metric Space Embedding by Unbalanced OT with Gromov-Wasserstein Marginal Penalization

    Florian Beier et al. “Joint Metric Space Embedding by Unbalanced OT with Gromov-Wasserstein Marginal Penalization”. In: Forty-second International Con- ference on Machine Learning (ICML) . 2025. url: https : / / openreview . net / forum?id=0YZHfUmsJv

    work page 2025

  6. [5]

    The Gene Mover's Distance: Single-cell similarity via Optimal Transport

    Riccardo Bellazzi et al. “The Gene Mover’s Distance: Single-cell similarity via Optimal Transport”. 2021. doi: 10.48550/arXiv.2102.01218

    work page Pith review arXiv doi:10.48550/arxiv.2102.01218 2021

  7. [6]

    Metric Learning

    Aur´ elien Bellet, Amaury Habrard, and Marc Sebban. Metric Learning . Vol. 9. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, 2015. doi: 10.2200/S00626ED1V01Y201501AIM030

    work page doi:10.2200/s00626ed1v01y201501aim030 2015

  8. [7]

    Geometric Analysis and Metric Learning of Instruction Em- beddings

    Sajib Biswas et al. “Geometric Analysis and Metric Learning of Instruction Em- beddings”. In: 2022 International Joint Conference on Neural Networks (IJCNN) . 2022, pp. 1–8. doi: 10.1109/IJCNN55064.2022.9892426

    arXiv 2022

Show all 68 references
  1. [8]

    Theory of Retracts

    Karol Borsuk. Theory of Retracts. Vol. 44. Monografie Matematyczne. Warszawa: Pa´ nstwowe Wydawnictwo Naukowe, 1967.isbn: 978-0800220815

    1967

  2. [9]

    ¨Uber Abbildung von Mannigfaltigkeiten

    L.E.J. Brouwer. “ ¨Uber Abbildung von Mannigfaltigkeiten”. In:Mathematische An- nalen 71 (1911), pp. 97–115. doi: 10.1007/BF01456931

    1911 doi

  3. [10]

    A Simple Framework for Contrastive Learning of Visual Rep- resentations

    Ting Chen et al. “A Simple Framework for Contrastive Learning of Visual Rep- resentations”. In: Proceedings of the 37th International Conference on Machine Learning. Ed. by Hal Daum´ e III and Aarti Singh. Vol. 119. ICML’20. JMLR, 13–18 Jul 2020, pp. 1597–1607. url: https://pr...

    2020

  4. [11]

    Unbalanced optimal transport: Dynamic and Kantorovich formulations

    Lenaic Chizat et al. “Unbalanced optimal transport: Dynamic and Kantorovich formulations”. In: Journal of Functional Analysis 274.11 (2018), pp. 3090–3123. doi: 10.1016/j.jfa.2018.03.008

    2018 doi

  5. [12]

    Fan R. K. Chung. Spectral Graph Theory . Vol. 92. CBMS Regional Conference Series in Mathematics. American Mathematical Society, 1997. isbn: 978-0-8218- 0315-8. doi: 10.1090/cbms/092. 37

    1997 doi

  6. [13]

    A Metric Learning Approach to Misogyny Categoriza- tion

    Juan Manuel Coria et al. “A Metric Learning Approach to Misogyny Categoriza- tion”. In: Proceedings of the 5th Workshop on Representation Learning for NLP . Ed. by Spandana Gella et al. Online: Association for Computational Linguistics, July 2020, pp. 89–94. doi: 10.18653/v1/2...

    2020 doi

  7. [14]

    Human Motion Analysis with Deep Metric Learning

    Huseyin Coskun et al. “Human Motion Analysis with Deep Metric Learning”. In: Proceedings of the European Conference on Computer Vision (ECCV). Sept. 2018. doi: 10.1007/978-3-030-01264-9_41

    2018 doi

  8. [15]

    Sinkhorn Distances: Lightspeed Computation of Optimal Trans- port

    Marco Cuturi. “Sinkhorn Distances: Lightspeed Computation of Optimal Trans- port”. In: Advances in Neural Information Processing Systems . Vol. 26. Curran Associates, Inc., 2013, pp. 2292–2300. url: https://papers.nips.cc/paper_ files / paper / 2013 / hash / af21d0c97db2e27e13...

    2013

  9. [16]

    Ground Metric Learning

    Marco Cuturi and David Avis. “Ground Metric Learning”. In: Journal of Machine Learning Research 15.1 (2014), pp. 533–564. url: http :/ /jmlr . org/ papers/ v15/cuturi14a.html

    2014

  10. [17]

    Information-theoretic metric learning

    Jason V. Davis et al. “Information-theoretic metric learning”. In: Proceedings of the 24th International Conference on Machine Learning . ICML ’07. Corvalis, Oregon, USA: Association for Computing Machinery, 2007, pp. 209–216. doi: 10.1145/ 1273496.1273523

    2007

  11. [18]

    An Optimal Transport Approach to Deep Metric Learning (Student Abstract)

    Jason Xiaotian Dou, Lei Luo, and Raymond Mingrui Yang. “An Optimal Transport Approach to Deep Metric Learning (Student Abstract)”. In: Proceedings of the AAAI Conference on Artificial Intelligence . Vol. 36. 11. 2022, pp. 12935–12936. doi: 10.1609/aaai.v36i11.21604

    2022 doi

  12. [19]

    Fast Unsu- pervised Ground Metric Learning with Tree-Wasserstein Distance

    Kira Michaela D¨ usterwald, Samo Hromadka, and Makoto Yamada. “Fast Unsu- pervised Ground Metric Learning with Tree-Wasserstein Distance”. In: The Thir- teenth International Conference on Learning Representations . 2025. url: https: //openreview.net/forum?id=FBhKUXK7od

    2025

  13. [20]

    Interpolating between Optimal Transport and MMD using Sinkhorn Divergences

    Jean Feydy et al. “Interpolating between Optimal Transport and MMD using Sinkhorn Divergences”. In: Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS) . Vol. 89. PMLR, 2019, pp. 2681–

    2019

  14. [21]

    Elements of Dimensionality Reduction and Manifold Learning

    Benyamin Ghojogh et al. Elements of Dimensionality Reduction and Manifold Learning. New York, NY: Springer, 2023. doi: 10.1007/978-3-031-10602-6

    2023 doi

  15. [22]

    Spectral, probabilistic, and deep metric learning: Tuto- rial and survey

    Benyamin Ghojogh et al. “Spectral, probabilistic, and deep metric learning: Tuto- rial and survey”. 2022. doi: 10.48550/arXiv.2201.09267

    2022 doi arXiv

  16. [23]

    Fixed Point Theory

    Andrzej Granas, James Dugundji, et al. Fixed Point Theory . Vol. 14. New York, NY: Springer, 2003. doi: 10.1007/978-0-387-21593-8

    2003 doi

  17. [24]

    Dimensionality Reduction by Learning an Invariant Mapping

    R. Hadsell, S. Chopra, and Y. LeCun. “Dimensionality Reduction by Learning an Invariant Mapping”. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06) . Vol. 2. 2006, pp. 1735–1742. doi: 10.1109/CVPR.2006.100. 38

    2006 doi

  18. [25]

    Integrated analysis of multimodal single-cell data

    Yuhan Hao et al. “Integrated analysis of multimodal single-cell data”. In: Cell 184.13 (2021), 3573–3587.e29. issn: 0092-8674. doi: 10.1016/j.cell.2021.04. 048

    2021 doi

  19. [26]

    Ground metric learning on graphs

    Matthieu Heitz et al. “Ground metric learning on graphs”. In: Journal of Math- ematical Imaging and Vision 63 (2021), pp. 89–107. doi: 10.1007/s10851-020- 00996-z

    2021 doi

  20. [27]

    Random function iterations for consistent stochastic feasibility

    Neal Hermer, D Russell Luke, and Anja Sturm. “Random function iterations for consistent stochastic feasibility”. In: Numerical Functional Analysis and Optimiza- tion 40.4 (2019), pp. 386–420. doi: 10.1080/01630563.2018.1535507

    2019

  21. [28]

    Random function iterations for stochastic fixed point problems

    Neal Hermer, D. Russell Luke, and Anja Sturm. “Random function iterations for stochastic fixed point problems”. 2020. doi: 10.48550/arXiv.2007.06479

    2020 doi arXiv

  22. [29]

    Horn and Charles R

    Roger A. Horn and Charles R. Johnson. Matrix Analysis. 2nd. Cambridge Univer- sity Press, 2012. isbn: 9780521386326

    2012

  23. [30]

    Deep Metric Learning for Visual Track- ing

    Junlin Hu, Jiwen Lu, and Yap-Peng Tan. “Deep Metric Learning for Visual Track- ing”. In: IEEE Transactions on Circuits and Systems for Video Technology 26.11 (2016), pp. 2056–2068. doi: 10.1109/TCSVT.2015.2477936

    2016

  24. [31]

    Supervised Word Mover’s Distance

    Gao Huang et al. “Supervised Word Mover’s Distance”. In: Advances in Neural Information Processing Systems. Ed. by D. Lee et al. Vol. 29. Curran Associates, Inc., 2016. url: https://proceedings.neurips.cc/paper_files/paper/2016/ file/10c66082c124f8afe3df4886f5e516e0-Paper.pdf

    2016

  25. [32]

    Python package wsingular

    Geert-Jan Huizing. Python package wsingular. https://github.com/CSDUlm/ wsingular. version:0.1.7. 2022

    2022

  26. [34]

    Mining on Manifolds: Metric Learning without Labels

    Ahmet Iscen et al. “Mining on Manifolds: Metric Learning without Labels”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018, pp. 7642–7651. doi: 10.1109/CVPR.2018.00797

    2018

  27. [35]

    Deep Metric Learning: A Survey

    Mahmut Kaya and Hasan S ¸akir Bilge. “Deep Metric Learning: A Survey”. In: Symmetry 11.9 (2019), p. 1066. doi: 10.3390/sym11091066

    2019 doi

  28. [36]

    Metric Learning in Optimal Transport for Domain Adaptation

    Tanguy Kerdoncuff, R´ emi Emonet, and Marc Sebban. “Metric Learning in Optimal Transport for Domain Adaptation”. In: International Joint Conference on Artifi- cial Intelligence . IJCAI. 2020, pp. 2162–2168. doi: 10.24963/ijcai.2020/299

    2020 doi

  29. [37]

    Unsupervised Meta- Learning for Few-Shot Image Classification

    Siavash Khodadadeh, Ladislau Boloni, and Mubarak Shah. “Unsupervised Meta- Learning for Few-Shot Image Classification”. In: Advances in Neural Information Processing Systems. Ed. by H. Wallach et al. Vol. 32. Curran Associates, Inc.,

  30. [38]

    Transfer operators from optimal transport plans for coherent set detection

    Peter Koltai et al. “Transfer operators from optimal transport plans for coherent set detection”. In: Physica D 426 (2021), p. 132980. doi: 10.1016/j.physd.2021. 132980

    2021 doi

  31. [39]

    Lahiri Krishna B

    Soumendra N. Lahiri Krishna B. Athreya. Measure Theory and Probability Theory. 1st ed. New York, NY: Springer New York, NY, 2006. doi: 10.1007/978-0-387- 35434-7

    2006 doi

  32. [40]

    From Word Embeddings To Document Distances

    Matt Kusner et al. “From Word Embeddings To Document Distances”. In: Pro- ceedings of the 32nd International Conference on Machine Learning. Ed. by Francis Bach and David Blei. Vol. 37. Proceedings of Machine Learning Research. Lille, France: PMLR, July 2015, pp. 957–966. url:...

    2015

  33. [41]

    Coupled Hierarchical Structure Learning using Tree- Wasserstein Distance

    Ya-Wei Eileen Lin et al. “Coupled Hierarchical Structure Learning using Tree- Wasserstein Distance”. 2025. doi: 10.48550/arXiv.2501.03627

    2025 doi

  34. [42]

    Hierarchical Neighbors Embedding

    Shenglan Liu et al. “Hierarchical Neighbors Embedding”. In: IEEE Transactions on Neural Networks and Learning Systems 35.6 (2024), pp. 7816–7829. doi: 10. 1109/TNNLS.2022.3221103

    2024

  35. [43]

    SphereFace: Deep Hypersphere Embedding for Face Recog- nition

    Weiyang Liu et al. “SphereFace: Deep Hypersphere Embedding for Face Recog- nition”. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2017, pp. 6738–6746. doi: 10.1109/CVPR.2017.713

    2017 doi

  36. [44]

    BioEncoder: A met- ric learning toolkit for comparative organismal biology

    Moritz D. L¨ urig, Emanuela Di Martino, and Arthur Porto. “BioEncoder: A met- ric learning toolkit for comparative organismal biology”. In: Ecology Letters 27.8 (2024). doi: 10.1111/ele.14495

    2024 doi

  37. [45]

    Visualizing Data using t-SNE

    Laurens van der Maaten and Geoffrey Hinton. “Visualizing Data using t-SNE”. In: Journal of Machine Learning Research 9.86 (2008), pp. 2579–2605. url: http: //jmlr.org/papers/v9/vandermaaten08a.html

    2008

  38. [46]

    Metric learning on expression data for gene function prediction

    Stavros Makrodimitris, Marcel J T Reinders, and Roeland C H J van Ham. “Metric learning on expression data for gene function prediction”. In: Bioinformatics 36.4 (Sept. 2019), pp. 1182–1190. issn: 1367-4803. doi: 10 . 1093 / bioinformatics / btz731

    2019

  39. [47]

    Interpolation of scattered data: distance matrices and condition- ally positive definite functions

    C. A. Micchelli. “Interpolation of scattered data: distance matrices and condition- ally positive definite functions”. In: Constructive Approximation 2 (1986), pp. 11–

    1986

  40. [48]

    DiV A: Diverse Visual Feature Aggregation for Deep Metric Learning

    Timo Milbich et al. “DiV A: Diverse Visual Feature Aggregation for Deep Metric Learning”. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII 16 . Springer. 2020, pp. 590–607. doi: 10.1007/978-3-030-58598-3_35

    2020 doi

  41. [49]

    Contributions to the problem of approximation of equidistant data by analytic functions

    I.J Schoenberg. “Contributions to the problem of approximation of equidistant data by analytic functions”. In:Quarterly of Applied Mathematics 4 (1946), pp. 45– 99. 40

    1946

  42. [50]

    doi: 10.1007/BF01893414

    doi

  43. [51]

    FaceNet: A Unified Embedding for Face Recognition and Clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. “FaceNet: A Unified Embedding for Face Recognition and Clustering”. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . June 2015, pp. 815–823. doi: 10.1109/CVPR.2015.7298682

    2015

  44. [52]

    Computational and analytical challenges in single-cell transcriptomics

    Oliver Stegle, Sarah A. Teichmann, and John C. Marioni. “Computational and analytical challenges in single-cell transcriptomics”. In: Nature Reviews Genetics 16.3 (2015), pp. 133–145. doi: 10.1038/nrg3833

    2015 doi

  45. [53]

    Bernhard Sch¨ olkopf and Alexander J. Smola. Learning with Kernels: Support Vec- tor Machines, Regularization, Optimization, and Beyond . Cambridge, MA, USA: MIT Press, 2001. isbn: 0262194759

    2001

  46. [54]

    A tutorial on dis- tance metric learning: Mathematical foundations, algorithms, experimental analy- sis, prospects and challenges

    Juan Luis Su´ arez, Salvador Garc ´ ıa, and Francisco Herrera. “A tutorial on dis- tance metric learning: Mathematical foundations, algorithms, experimental analy- sis, prospects and challenges”. In: Neurocomputing 425 (2021), pp. 300–322. doi: 10.1016/j.neucom.2020.08.017

    2021 doi

  47. [55]

    Deep Metric Learning: a (Long) Survey

    Chan Ha Vu. Deep Metric Learning: a (Long) Survey. 2021. url: https://hav4ik. github.io/articles/deep-metric-learning-survey

    2021

  48. [56]

    Inverse Optimal Transport

    Andrew M. Stuart and Marie-Therese Wolfram. “Inverse Optimal Transport”. In: SIAM Journal on Applied Mathematics 80.1 (2020), pp. 257–279. doi: 10.1137/ 19M1261122

    2020

  49. [57]

    CosFace: Large Margin Cosine Loss for Deep Face Recognition

    Hao Wang et al. “CosFace: Large Margin Cosine Loss for Deep Face Recognition”. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2018, pp. 5265–5274. doi: 10.1109/CVPR.2018.00552

    2018

  50. [58]

    Distance Metric Learning for Large Margin Nearest Neighbor Classification

    Kilian Q Weinberger, John Blitzer, and Lawrence Saul. “Distance Metric Learning for Large Margin Nearest Neighbor Classification”. In: Advances in Neural Infor- mation Processing Systems 18 (2005). url: https : / / proceedings . neurips . cc / paper _ files / paper / 2005 / fi...

    2005

  51. [59]

    Supervised Earth Mover’s Distance Learning and Its Computer Vision Applications

    Fan Wang and Leonidas J Guibas. “Supervised Earth Mover’s Distance Learning and Its Computer Vision Applications”. In: Computer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12 . Springer, 2012, pp. 442–...

    2012 doi

  52. [60]

    SCANPY: Large-scale single-cell gene expression data analysis

    F. Wolf, Philipp Angerer, and Fabian Theis. “SCANPY: Large-scale single-cell gene expression data analysis”. In: Genome Biology 19 (Feb. 2018). doi: 10.1186/ s13059-017-1382-0 . 41

    2018

  53. [61]

    Distance Metric Learning with Application to Clustering with Side-Information

    Eric Xing et al. “Distance Metric Learning with Application to Clustering with Side-Information”. In: Advances in Neural Information Processing Systems . Ed. by S. Becker, S. Thrun, and K. Obermayer. Vol. 15. 2002, pp. 505–512. url: https://papers.nips.cc/paper/2164- distance-...

    2002

  54. [62]

    Scattered Data Approximation

    Holger Wendland. Scattered Data Approximation. Cambridge Monographs on Ap- plied and Computational Mathematics. Cambridge University Press, 2005. doi: 10.1017/CBO9780511617539

    2005 doi

  55. [63]

    Spectral Relaxation for K-means Clustering

    Hongyuan Zha et al. “Spectral Relaxation for K-means Clustering”. In: Ad- vances in Neural Information Processing Systems (NeurIPS) . Ed. by T. Diet- terich, S. Becker, and Z. Ghahramani. Vol. 14. MIT Press, 2001, pp. 1057–1064. url: https : / / proceedings . neurips . cc / pa...

    2001

  56. [65]

    Multi-Level Metric Learning via Smoothed Wasserstein Distance

    Jie Xu et al. “Multi-Level Metric Learning via Smoothed Wasserstein Distance”. In: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI) . 2018, pp. 2919–2925. doi: 10.24963/ijcai.2018/405

    2018 doi

  57. [67]

    contractible, if there exists a pointaP M and deformation retract onto the singleton tau

  58. [68]

    Now, we verify that all conditions are satisfied

    (strongly) locally contractible , if for every point a P M and every neighborhood VĂ M of a there exists a neighborhood UĂ V that is contractible in V. Now, we verify that all conditions are satisfied. Compactness of Mr m follows directly from its definition. We can view Mr m ...

  59. [2019]

    url: https : / / proceedings . neurips . cc / paper _ files / paper / 2019 / file/fd0a5a5e367a0955d81278062ef37429-Paper.pdf. 39

    2019

  60. [2690]

    url: https://proceedings.mlr.press/v89/feydy19a.html

Pith tools

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