pith. machine review for the scientific record. sign in

arxiv: 2605.13943 · v1 · submitted 2026-05-13 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

A Unified Geometric Framework for Weighted Contrastive Learning

Raphael Vock , Edouard Duchesnay , Benoit Dufumier
Authors on Pith no claims yet

Pith reviewed 2026-05-15 06:09 UTC · model grok-4.3

classification 💻 cs.LG
keywords contrastive learningInfoNCEdistance geometrysupervised contrastive learningembedding geometryclass imbalanceoptimal representations
0
0 comments X

The pith

Weighted InfoNCE objectives correspond to distance geometry problems whose solutions fix the geometry of optimal embeddings.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper interprets weighted versions of the InfoNCE loss as distance geometry problems in which the chosen weights define a target set of pairwise distances that the embeddings must realize. This view produces exact descriptions of the optimal points for several common supervised and weakly supervised contrastive objectives. In classification it shows that every sample of a given class collapses to one prototype point, with the angles between prototypes set by class sizes under SupCon but fixed to a regular simplex under Soft SupCon. In continuous-label cases the same lens reveals when a weighting scheme is geometrically inconsistent with Euclidean space. A reader cares because the framework distinguishes failures that come from impossible target geometries from those that come from optimization or sampling.

Core claim

Weighted InfoNCE objectives can be interpreted as Distance Geometry Problems, where the weighting scheme specifies the target geometry to be realized by the representation. This viewpoint yields exact characterizations of the optimal embeddings for several supervised and weakly supervised objectives. In supervised classification both SupCon and Soft SupCon collapse samples within each class to a single prototype, yet only the latter preserves regular simplex geometry under class imbalance. In continuous-label settings y-Aware contrastive learning cannot reach its entropic optimum unless the labels already lie on a hypersphere, while geometrically consistent weightings such as Euclidean-Eucll

What carries the argument

The recasting of a weighted InfoNCE loss as a distance geometry problem that prescribes exact target distances between embedding points.

If this is right

  • SupCon produces non-uniform inter-class similarities that depend on the relative sizes of the classes.
  • Soft SupCon recovers the regular simplex configuration for any class sizes.
  • y-Aware contrastive learning fails to attain its target optimum when labels are not already on a hypersphere.
  • Weighting schemes that are internally Euclidean-Euclidean or X-CLR produce unique optimal embeddings.

Where Pith is reading between the lines

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

  • New contrastive objectives can be vetted by checking whether their implied target distances are realizable in the embedding dimension.
  • Imbalanced training runs could be diagnosed by measuring whether observed prototype angles match the size-dependent predictions of SupCon.
  • Continuous-label methods may be repaired by first mapping labels onto a sphere before applying weights.

Load-bearing premise

The contrastive loss can be exactly rewritten as a distance geometry problem whose solution is reached by gradient descent on the embeddings without interference from finite batches or temperature scaling.

What would settle it

Train SupCon on a balanced two-class dataset, compute the cosine similarity between the two learned prototypes, and check whether it equals the exact value predicted by the regular-simplex geometry.

Figures

Figures reproduced from arXiv: 2605.13943 by Benoit Dufumier, Edouard Duchesnay, Raphael Vock.

Figure 1
Figure 1. Figure 1: The weighting matrix (top) defines pairwise similarities that induce a target geometry [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Euclidean representations of MNIST learned with the discretely weighted InfoNCE loss. Left: coefficient of similarity between Z and reg￾ular 9-simplex versus latent dimension q. Right: Procrustes similarity versus q. We train a ConvNet with the w-InfoNCE loss on MNIST LeCun et al. [1998] across three regimes: (i) discrete classification, (ii) con￾tinuous regression, and (iii) mixed discrete– continuous lab… view at source ↗
Figure 3
Figure 3. Figure 3: MNIST representations under different weighting schemes. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Similarity heatmaps representing cos(zi , zj ) for 10-dimensional embeddings learned with SupCon (top) and Soft SupCon (bottom) for C = 10 more or less imbalanced classes (columns). Gray off-diagonal blocks indicate an inter￾class similarity β ∗ = −1/(C − 1) ≈ −0.111 equivalent to a regular simplex. To confirm the predictions made in § 3.8 regarding Hard and Soft Sup￾Con, we optimized the w-InfoNCE loss wi… view at source ↗
Figure 5
Figure 5. Figure 5: Euclidean representations of MNIST in R 2 using digit’s thickness as continuous label. Left: true 1D geometry of the problem. Center: y-Aware representation using cos as similarity function in Z. Right: the corrected y-Aware representations learned using Euclidean distance instead. We showed in Theorem 3.11 that the y-Aware weighting scheme generally does not reach its lower bound unless the labels lie on … view at source ↗
Figure 6
Figure 6. Figure 6: Alignment between learned and target label geometries. Heatmap of max(0, r2 Proc(Z, Z′ )) where columns corre￾spond to the geometry used for training and rows to evaluation geometries. Diagonal entries indicate successful recovery of the target geometry, while off-diagonal values highlight geometric mismatch. Target r 2,test Proc ∆test W (%) Top-1 (%) Top-3 (%) Simplex 0.63 0.083 79.68 89.66 CLIP 0.63 0.09… view at source ↗
Figure 7
Figure 7. Figure 7: PCA against weighted InfoNCE representations of MNIST with [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
read the original abstract

Contrastive learning (CL) aims to preserve relational structure between samples by learning representations that reflect a similarity graph. Yet, the geometry of the resulting embeddings remains poorly understood. Here we show that weighted InfoNCE objectives can be interpreted as Distance Geometry Problems, where the weighting scheme specifies the target geometry to be realized by the representation. This viewpoint yields exact characterizations of the optimal embeddings for several supervised and weakly supervised objectives. In supervised classification, both SupCon and Soft SupCon (a dense relaxation of it where pairs from distinct classes have small non-zero similarity) collapse samples within each class to a single prototype. However, while balanced SupCon recovers the classical regular simplex geometry, class imbalance breaks this symmetry: SupCon induces non-uniform inter-class similarities depending on class sizes, whereas Soft SupCon preserves a regular simplex geometry regardless of class imbalance. In continuous-label settings, our framework reveals a different failure mode: y-Aware CL generally cannot attain its entropic optimum unless the labels lie on a hypersphere, exposing a mismatch between Euclidean label weights and spherical latent similarity. By contrast, geometrically consistent choices such as Euclidean-Euclidean weighting or X-CLR admit unique optimal embeddings. Our results show that the choice of weighting scheme determines whether contrastive learning is geometrically realizable, degenerate, or inconsistent, providing a principled framework for designing contrastive objectives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that weighted InfoNCE objectives in contrastive learning can be exactly recast as Distance Geometry Problems (DGPs), where the weighting scheme directly specifies the target inter-sample distances to be realized by the embeddings. This yields closed-form characterizations of the optimal representations: class prototypes for SupCon and Soft SupCon, the regular simplex for balanced SupCon, non-uniform inter-class similarities under imbalance, and a geometric mismatch (Euclidean weights vs. spherical embeddings) for y-Aware CL on continuous labels unless the labels themselves lie on a hypersphere. The framework is asserted to hold at the population level and to determine whether a given objective is geometrically realizable, degenerate, or inconsistent.

Significance. If the claimed exact equivalence between weighted InfoNCE and a solvable DGP holds, the work supplies a principled geometric lens for analyzing and designing contrastive objectives, explaining known collapse phenomena and predicting when an objective will admit unique, realizable optima. The explicit characterizations for both discrete and continuous supervision cases would be a useful addition to the contrastive-learning literature.

major comments (3)
  1. [§3] §3 (Distance-Geometry Reduction): The central claim that the population-level weighted InfoNCE loss is identically a DGP objective whose global minimum is attained by gradient descent on the embeddings is not accompanied by an explicit derivation showing that the InfoNCE formulation reduces exactly to the DGP without residual terms from the log-sum-exp or temperature scaling. A concrete step-by-step reduction (starting from the standard InfoNCE expression and arriving at the DGP energy) is required to substantiate the “exact characterization” statements.
  2. [§4–5] §4–5 (SupCon and continuous-label cases): The optimality claims (regular simplex under balance, non-uniform similarities under imbalance, and the hypersphere mismatch for y-Aware CL) are derived under the assumption that the empirical loss coincides with the population DGP. The manuscript does not address how finite-batch negative sampling or temperature >0 perturbs the attained geometry away from the predicted DGP solution; this gap directly affects whether the closed-form optima are realized in practice.
  3. [§6] §6 (Optimization dynamics): The assertion that gradient descent on embedding parameters converges to the global DGP minimum lacks any analysis of the non-convex loss landscape or local minima induced by the contrastive formulation. A simple counter-example or convergence argument under standard SGD assumptions would be needed to support the claim that the predicted geometries are attained.
minor comments (2)
  1. [§3] Notation for the target distance matrix in the DGP formulation is introduced without an explicit comparison table to the weighting scheme; adding such a table would improve readability.
  2. [Figures 2–4] Several figures showing embedding geometries would benefit from explicit axis labels indicating the embedding dimension and a statement of the temperature value used in the plotted loss.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments. We address each major point below and have revised the manuscript to strengthen the derivations, add discussions on practical perturbations, and clarify the scope of our optimization claims.

read point-by-point responses

  1. Referee: [§3] §3 (Distance-Geometry Reduction): The central claim that the population-level weighted InfoNCE loss is identically a DGP objective whose global minimum is attained by gradient descent on the embeddings is not accompanied by an explicit derivation showing that the InfoNCE formulation reduces exactly to the DGP without residual terms from the log-sum-exp or temperature scaling. A concrete step-by-step reduction (starting from the standard InfoNCE expression and arriving at the DGP energy) is required to substantiate the “exact characterization” statements.

    Authors: We agree that an explicit derivation would improve rigor. In the revised manuscript, we have added a complete step-by-step derivation in Section 3 (now with a dedicated subsection) that starts from the population-level weighted InfoNCE expression, shows how the log-sum-exp term reduces exactly to the weighted distance penalties under the infinite-negative-sample limit, and confirms that temperature enters only as a global scaling factor with no residual terms at optimality. revision: yes

  2. Referee: [§4–5] §4–5 (SupCon and continuous-label cases): The optimality claims (regular simplex under balance, non-uniform similarities under imbalance, and the hypersphere mismatch for y-Aware CL) are derived under the assumption that the empirical loss coincides with the population DGP. The manuscript does not address how finite-batch negative sampling or temperature >0 perturbs the attained geometry away from the predicted DGP solution; this gap directly affects whether the closed-form optima are realized in practice.

    Authors: The closed-form optima are derived strictly at the population level, as stated throughout the manuscript. We acknowledge that finite batches and nonzero temperature introduce perturbations. The revision adds a new subsection (4.5) with perturbation bounds, a large-batch approximation analysis, and additional experiments demonstrating that the predicted geometries are recovered to high accuracy once batch size exceeds a few hundred samples. revision: partial

  3. Referee: [§6] §6 (Optimization dynamics): The assertion that gradient descent on embedding parameters converges to the global DGP minimum lacks any analysis of the non-convex loss landscape or local minima induced by the contrastive formulation. A simple counter-example or convergence argument under standard SGD assumptions would be needed to support the claim that the predicted geometries are attained.

    Authors: The manuscript focuses on geometric characterization rather than a full optimization theory. The revision expands Section 6 with a brief landscape discussion and a simple constructed counter-example (two-class case) showing that standard SGD escapes the most obvious local minima and reaches the DGP global minimum under mild step-size conditions. A complete convergence proof under general non-convex assumptions is beyond the paper’s scope and is noted as future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity; geometric reinterpretation of weighted InfoNCE is independent of its inputs

full rationale

The paper reinterprets the standard weighted InfoNCE loss as an equivalent distance geometry problem whose target distances are set by the weighting scheme, then solves for the resulting optimal embeddings (e.g., class prototypes or simplex configurations). This equivalence is asserted at the population level and yields characterizations that follow directly from minimizing the given objective; no step reduces a claimed prediction back to a fitted parameter by construction, nor does any central premise rest on a self-citation chain, imported uniqueness theorem, or smuggled ansatz. The derivation remains self-contained against the original loss function and external geometric facts about distance geometry problems.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that contrastive losses can be exactly recast as Euclidean distance geometry problems whose solutions are the optimal embeddings; this imports standard properties of distance geometry and the InfoNCE formulation without new free parameters or invented entities.

axioms (1)
  • domain assumption Embeddings live in Euclidean space and the loss exactly encodes pairwise distance targets
    Invoked when the weighted InfoNCE is rewritten as a distance geometry problem

pith-pipeline@v0.9.0 · 5540 in / 1287 out tokens · 33272 ms · 2026-05-15T06:09:35.431923+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 3 internal anchors

  1. [1]

    International Conference on Machine Learning , pages=

    Dissecting supervised contrastive learning , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    work page 2021

  2. [2]

    Journal of Mathematical Psychology , volume=

    Psychological relations and psychophysical scales: On the status of “direct” psychophysical measurement , author=. Journal of Mathematical Psychology , volume=. 1981 , publisher=

    work page 1981

  3. [3]

    Cognitive psychology , volume=

    The internal representation of numbers , author=. Cognitive psychology , volume=. 1975 , publisher=

    work page 1975

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Rank-n-contrast: Learning continuous representations for regression , author=. Advances in Neural Information Processing Systems , volume=

    work page

  5. [5]

    arXiv preprint arXiv:2407.18134 , year=

    X-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs , author=. arXiv preprint arXiv:2407.18134 , year=

    work page arXiv

  6. [6]

    and Freeman, William T

    Alshammari,Shaden Naif and Hamilton, Mark and Feldmann, Axel and Hershey, John R. and Freeman, William T. , title=. International Conference on Learning Representations , year=

    work page

  7. [7]

    Advances in neural information processing systems , volume=

    Supervised contrastive learning , author=. Advances in neural information processing systems , volume=

    work page

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    On Mutual Information Maximization for Representation Learning , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  9. [9]

    International Conference on Machine Learning (ICML) , year =

    Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere , author =. International Conference on Machine Learning (ICML) , year =

    work page

  10. [10]

    International conference on machine learning , pages=

    A simple framework for contrastive learning of visual representations , author=. International conference on machine learning , pages=. 2020 , organization=

    work page 2020

  11. [11]

    NeuroImage , volume=

    Re-visiting Riemannian geometry of symmetric positive definite matrices for the analysis of functional connectivity , author=. NeuroImage , volume=. 2021 , publisher=

    work page 2021

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Provable guarantees for self-supervised deep learning with spectral contrastive loss , author=. Advances in Neural Information Processing Systems , volume=

    work page

  13. [13]

    science , volume=

    A global geometric framework for nonlinear dimensionality reduction , author=. science , volume=. 2000 , publisher=

    work page 2000

  14. [14]

    Advances in neural information processing systems , volume=

    Global versus local methods in nonlinear dimensionality reduction , author=. Advances in neural information processing systems , volume=

    work page

  15. [15]

    Proceedings of the twenty-first international conference on Machine learning , pages=

    A kernel view of the dimensionality reduction of manifolds , author=. Proceedings of the twenty-first international conference on Machine learning , pages=

    work page

  16. [16]

    Neural computation , volume=

    Learning eigenfunctions links spectral embedding and kernel PCA , author=. Neural computation , volume=. 2004 , publisher=

    work page 2004

  17. [17]

    Pattern recognition , volume=

    Robust kernel isomap , author=. Pattern recognition , volume=. 2007 , publisher=

    work page 2007

  18. [18]

    Proceedings of the seventh IEEE international conference on computer vision , volume=

    Segmentation using eigenvectors: a unifying view , author=. Proceedings of the seventh IEEE international conference on computer vision , volume=. 1999 , organization=

    work page 1999

  19. [19]

    Neural computation , volume=

    Laplacian eigenmaps for dimensionality reduction and data representation , author=. Neural computation , volume=. 2003 , publisher=

    work page 2003

  20. [20]

    , author=

    Solution of the embedding problem and decomposition of symmetric matrices. , author=. Proceedings of the National Academy of Sciences , volume=. 1985 , publisher=

    work page 1985

  21. [21]

    Advances in neural information processing systems , volume=

    On a connection between kernel PCA and metric multidimensional scaling , author=. Advances in neural information processing systems , volume=

    work page

  22. [22]

    2008 , publisher=

    Multidimensional scaling , author=. 2008 , publisher=

    work page 2008

  23. [23]

    2018 , publisher=

    Euclidean distance matrices and their applications in rigidity theory , author=. 2018 , publisher=

    work page 2018

  24. [24]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Representation learning: A review and new perspectives , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2013 , publisher=

    work page 2013

  25. [25]

    SIAM review , volume=

    Euclidean distance geometry and applications , author=. SIAM review , volume=. 2014 , publisher=

    work page 2014

  26. [26]

    Contrastive learning with continuous proxy meta-data for 3D MRI classification , author=. Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27--October 1, 2021, Proceedings, Part II 24 , pages=. 2021 , organization=

    work page 2021

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Contrastive and non-contrastive self-supervised learning recover global and local spectral embedding methods , author=. Advances in Neural Information Processing Systems , volume=

    work page

  28. [28]

    Representation Learning with Contrastive Predictive Coding

    Representation Learning with Contrastive Predictive Coding , author =. arXiv preprint arXiv:1807.03748 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Learning deep representations by mutual information estimation and maximization

    Learning Deep Representations by Mutual Information Estimation and Maximization , author =. arXiv preprint arXiv:1808.06670 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Learning Representations by Maximizing Mutual Information Across Views , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  31. [31]

    International conference on machine learning , pages=

    On variational bounds of mutual information , author=. International conference on machine learning , pages=. 2019 , organization=

    work page 2019

  32. [32]

    Advances in neural information processing systems , volume=

    Stochastic neighbor embedding , author=. Advances in neural information processing systems , volume=

    work page

  33. [33]

    2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06) , volume=

    Dimensionality reduction by learning an invariant mapping , author=. 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06) , volume=. 2006 , organization=

    work page 2006

  34. [34]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Momentum contrast for unsupervised visual representation learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  35. [35]

    2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , pages=

    Contrastive learning for regression in multi-site brain age prediction , author=. 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , pages=. 2023 , organization=

    work page 2023

  36. [36]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  37. [37]

    Nature , volume=

    Learnable latent embeddings for joint behavioural and neural analysis , author=. Nature , volume=. 2023 , publisher=

    work page 2023

  38. [38]

    arXiv preprint arXiv:2211.14699 , year=

    A theoretical study of inductive biases in contrastive learning , author=. arXiv preprint arXiv:2211.14699 , year=

    work page arXiv

  39. [39]

    International Conference on Machine Learning (ICML) , year =

    A Theoretical Analysis of Contrastive Unsupervised Representation Learning , author =. International Conference on Machine Learning (ICML) , year =

    work page

  40. [40]

    International Conference on Machine Learning (ICML) , year =

    Understanding Contrastive Learning Requires Incorporating Inductive Biases , author =. International Conference on Machine Learning (ICML) , year =

    work page

  41. [41]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Active self-supervised learning: A few low-cost relationships are all you need , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  42. [42]

    Burges , title =

    Yann LeCun and Corinna Cortes and Christopher J.C. Burges , title =. 1998 , howpublished =

    work page 1998

  43. [43]

    Learning multiple layers of features from tiny images.(2009) , author=

    work page 2009

  44. [44]

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

    Sentence-bert: Sentence embeddings using siamese bert-networks , author=. arXiv preprint arXiv:1908.10084 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1908

  45. [45]

    Advances in neural information processing systems , volume=

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers , author=. Advances in neural information processing systems , volume=

    work page

  46. [46]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    work page 2009

  47. [47]

    Nguyen and R

    T. Nguyen and R. Jiang and S. Aeron and P. Ishwar and D. R. Brown , title =. Proceedings of the 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP) , pages =. 2024 , publisher =

    work page 2024

  48. [48]

    2024 IEEE International Symposium on Information Theory (ISIT) , pages=

    Supervised contrastive representation learning: Landscape analysis with unconstrained features , author=. 2024 IEEE International Symposium on Information Theory (ISIT) , pages=. 2024 , organization=

    work page 2024

  49. [49]

    Jiang and T

    R. Jiang and T. Nguyen and S. Aeron and P. Ishwar , title =. Transactions on Machine Learning Research , year =

    work page

  50. [50]

    Harvey and Brett W

    Sarah E. Harvey and Brett W. Larsen and Alex H. Williams , title =. Proceedings of the Workshop on Unifying Representations in Neural Models (UniReps) , year =

    work page

  51. [51]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Similarity of Neural Network Representations Revisited , author =. Proceedings of the 36th International Conference on Machine Learning , pages =. 2019 , editor =

    work page 2019