Pith. sign in

REVIEW 4 major objections 4 minor 121 references

Constrained Sliced Wasserstein Embedding

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Constrained transport plans sharpen sliced Wasserstein pooling.

desk verdict Solid incremental contribution to sliced Wasserstein embedding, but the claimed mechanism (SWGG constraints improving slices) is under-verified because constraint satisfaction is never checked on the hard sorting used at inference. read the letter →

arxiv 2506.02203 v1 pith:X2ZPG47U submitted 2025-06-02 cs.LG cs.AImath.OCq-bio.QMstat.ML

classification cs.LGcs.AImath.OCq-bio.QMstat.ML
keywords slicedWassersteindistanceoptimaltransportconstrainedlearningprimal-dualoptimizationpermutation-invariantpoolingembeddingsoftsortpointcloudclassification
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

Sliced Wasserstein embedding compares high-dimensional sets of vectors by projecting them onto one-dimensional lines and recording how mass moves along each line. The paper's claim is that the lines matter: if each line's one-dimensional transport plan is forced, through a constraint, to stay close to the optimal transport plan in the original space, the resulting slices are more informative, and downstream classifiers need fewer slices to reach a given accuracy. The paper implements the constraint with the sliced Wasserstein generalized geodesic dissimilarity $D_2(\mu,\nu;\theta)$, relaxes it with slack variables to preserve feasibility, and trains the whole pooling layer with a primal-dual algorithm. In experiments on image, point cloud, and protein-sequence classification, the constrained embedding matches or beats the unconstrained one at equal slice budgets, with the largest gains where slices are scarce.

What carries the argument

The load-bearing object is the sliced Wasserstein generalized geodesic dissimilarity $D_2(\mu,\nu;\theta)$, built from the lifted transport plan $\gamma^{\mu,\nu}_\theta$: one takes the unique 1D optimal transport plan between $\theta_\#\mu$ and $\theta_\#\nu$, lifts it to the original space with the weights $u^{\mu,\nu}_\theta(x,y) = p(x)q(y)\Lambda^{\mu,\nu}_\theta(\theta^T x,\theta^T y)/(P_\theta(x)Q_\theta(y))$, and computes the cost of that lifted plan. Because $W_2(\mu,\nu)\le D_2(\mu,\nu;\theta)$ for every direction, a slice with small $D_2$ is, by construction, one whose 1D plan is close to globally optimal. The paper turns this inequality into constraints, relaxes the constraints with slack variables, and differentiates through the sorting-based permutation matrices with the softsort operator, a continuous relaxation of the argsort permutation, so the entire pooling layer is trainable by primal-dual gradient updates.

What would settle it

A direct test would compute, on a held-out split of any of the three tasks, the correlation between per-slice $D_2(\mu,\nu;\theta_l)$ and that slice's contribution to classification accuracy; the paper's premise predicts a positive correlation, and a null or negative correlation would settle against it. A sharper synthetic version is a binary classification problem whose class difference is carried by a high-frequency mode that the optimal transport plan ignores; if constrained SWE satisfies its bounds yet underperforms unconstrained SWE there, the small-$D_2$ premise is false.

Watch

Extended reading notes

Core claim

The central claim is that constraining a slice's lifted transport plan to approximate the optimal plan in the original space makes sliced Wasserstein embeddings learn better slicing directions. For a pair of measures $\mu$ and $\nu$, the paper imposes $D_2(\mu,\nu;\theta_l)\le \epsilon_l+s_l$ on each slice, where $D_2$ is the SWGG dissimilarity built by lifting the 1D optimal transport plan back into $\mathbb{R}^d$, and $W_2(\mu,\nu)\le D_2(\mu,\nu;\theta)$ holds for every direction. Slack variables $s_l$ keep the problem feasible when the bounds are tight, and the dual variables in the primal-dual update automatically up-weight slices that violate their bounds, so training amounts to an adaptive regularization toward globally meaningful transport. Empirically, this yields consistent accuracy gains over unconstrained SWE on point cloud classification at every tested slice count, reduces overfitting on the image task, and gives the largest gains at small slice counts on protein subcellular localization.

Load-bearing premise

The load-bearing premise is the hypothesis that a slicing direction is informative for the downstream task precisely when moving mass along that direction costs nearly as much as moving it optimally in the full space; the paper states this as a hypothesis in Section 3 and supports it only with empirical results on three datasets.

Editorial extensions

If this is right

  • At a fixed slice budget, constrained SWE reaches higher classification accuracy than unconstrained SWE; on point cloud classification the gain holds for every tested number of slices, so the same accuracy can be reached with fewer slices and proportionally less compute.
  • Because the SWE embedding size grows linearly with the number of slices, more informative slices translate directly into smaller pooled representations and cheaper linear classification heads.
  • On the image task, constrained SWE keeps train accuracy far closer to validation accuracy than unconstrained SWE does, indicating that the SWGG constraint limits overfitting of the pooling layer.
  • On protein subcellular localization, a CLS token embedding matches roughly 16 constrained slices, and the advantage of the constraint is largest at $L=1$ and $L=4$; the advantage fades at larger $L$, consistent with the constrained problem becoming infeasible as more slices must share tight bounds.

Reading between the lines

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

  • The per-slice dual variables $\lambda_l$ measure how far each slice is from satisfying its SWGG bound, so they can be read as learned slice-importance weights; using them to weight or prune slices inside the final embedding is a natural follow-up that the paper does not test.
  • The constraint is defined entirely through the transport plan, not through the classification head, so the same primal-dual slice selection should apply to other sliced optimal transport objectives, such as sliced Wasserstein generative losses or barycenter problems; that transfer is outside the paper's experiments.
  • A direct test of the SWGG hypothesis would compare learned slice directions against the displacement directions of the original-space optimal transport plan; if constrained slices align more strongly with those displacements than unconstrained slices do, that would support the mechanism the paper proposes.
  • Because softsort is used only during training, inference retains the $O(M\log M)$ sorting cost; the method therefore scales to large token sets, with the temperature $\tau$ controlling a training-time bias whose effect on final accuracy the paper does not isolate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a constrained-learning formulation for sliced Wasserstein embeddings (SWE). It adds per-slice upper-bound constraints on the sliced-Wasserstein generalized geodesic (SWGG) dissimilarity D2(µ,ν;θ_l), introduces slack variables for feasibility, and solves the resulting problem with a primal-dual algorithm. To make the constraints differentiable, the authors replace the hard permutation matrices in the SWGG expression with a softsort relaxation in Eq. (23), while inference still uses hard argsort-based permutations in Eq. (18). The method is evaluated as a pooling layer for classification on Tiny ImageNet with a DeiT backbone, ModelNet40 point clouds with PCT, and protein subcellular localization with four ESM-2 models, reporting accuracy gains over unconstrained SWE and GAP/CLS baselines.

Significance. If the central mechanism were verified, the paper would make a useful contribution: a principled way to learn more informative slices for SW distances, potentially reducing the number of slices and the computational cost of sliced Wasserstein pooling. The empirical evaluation is internally consistent across three tasks with multiple seeds, and the code is released. The main gaps are that hard constraint satisfaction is never reported, comparisons to existing informative-slice baselines are missing, and the constraint is entangled with the trainable reference set U; these gaps currently prevent the paper from fully supporting the claim that the constraints, rather than auxiliary regularization effects, drive the observed improvements.

major comments (4)
  1. [Section 4.1, Eqs. (16), (21)-(23), and Appendix D] The constraint mechanism is trained exclusively through the softsort relaxation, but constraint satisfaction is never evaluated with the hard permutation matrices used at inference in Eq. (18). The paper should report hard SWGG dissimilarities D2(µl,νl_i;θ_l) for the final slices, the fraction of slices satisfying D2 ≤ ϵ_l + s_l with the trained slack variables, and the same quantities for unconstrained SWE. Without this, the accuracy gains in Section 5 cannot be attributed to the claimed constraints rather than to the softsort temperature or the slack/dual regularization. Appendix D is also ambiguous: it does not state whether the plotted SWGG values are computed with soft or hard permutations, and for ϵ = 5 the displayed average SWGG levels exceed 5, indicating that the slack variables are active.
  2. [Section 5, compared with Section 2.2] The empirical evaluation only compares constrained SWE with unconstrained SWE and GAP/CLS baselines; it does not compare against existing informative-slice methods such as Max-SW [24], Distributional-SW [79], Energy-Based SW [78], or Markovian SW [80], which are discussed in Section 2.2. Since the central claim is that the proposed method learns 'more informative slices,' the absence of these baselines leaves the claim unsupported relative to prior work. Adding at least one or two of these baselines, or explicitly explaining why they are not applicable to the SWE pooling setting, would be needed to substantiate the informativeness claim.
  3. [Section 4, Eq. (20)] The constraint D2(µl,νl_i;θ_l) depends not only on the slice θ_l but also on the trainable reference set U, which is a primal variable in Eq. (20). The optimization can therefore satisfy the SWGG upper bounds by modifying U (for example, by shrinking or relocating the reference embeddings) rather than by improving the slicing directions. This entanglement weakens the interpretation that the observed gains come from better slices. The authors should either ablate the effect by freezing U during constraint enforcement, or report diagnostics that separate the contribution of U from that of Θ to constraint satisfaction and downstream accuracy.
  4. [Section 3, the hypothesis preceding Eq. (10)] The paper states as a hypothesis that 'a good slice θ is one for which the SWGG dissimilarity D2(µ,ν;θ) is as small as possible,' but this premise is not derived or independently tested. The empirical gains in Sections 5.1-5.3 are consistent with the hypothesis, but they do not isolate it from other effects of the constrained training procedure, such as regularization of Θ or U. A direct test, such as comparing slices selected by minimizing D2 against slices selected by maximizing projected distance on the same downstream metric, or correlating per-slice D2 values with per-slice utility, would strengthen the paper's central claim.
minor comments (4)
  1. [Section 4.1, Eq. (23)] The SoftSort expression in Eq. (23) contains the non-differentiable sort(x) inside the softmax; the text should clarify that differentiability is with respect to the x-dependent term (or specify a straight-through treatment), so that the phrase 'differentiable approximation' is precise.
  2. [Appendix D, Figure 6] The caption of Figure 6 should state explicitly whether the plotted SWGG levels are computed with the soft permutation matrices used during training or with the hard permutation matrices used at inference; this is essential for interpreting the constraint-satisfaction behavior.
  3. [Section 5.2, Figure 3] The y-axis in Figure 3 begins at 0.75, which can visually exaggerate the differences between methods; consider starting the axis at 0 or adding a note about the truncated range.
  4. [Section 6, limitations paragraph] The limitations paragraph discusses future work on weighting slices with dual or slack variables, but it does not mention that the constraints in Eq. (20b) are enforced only on average over the training set; per-sample constraint satisfaction is therefore not guaranteed, which is relevant to the claim that the slices are meaningful in the original space.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SWGG constraint is an explicit design hypothesis, and the claimed accuracy gains are measured on external benchmarks independently of how the constraint is defined.

full rationale

The paper's derivation chain is not circular. The SWE pooling operator is taken from prior work [74]; the SWGG dissimilarity D2(mu,nu;theta) is defined in Eqs. (5)-(8) from [64,59]; the primal-dual constrained learning machinery is standard [14,15,37]; and the differentiable permutation relaxation is from softsort [87] with the formula stated in Eq. (23). The central claim is that imposing upper bounds on D2 during training improves downstream classification accuracy. That claim is evaluated by test accuracy on Tiny ImageNet, ModelNet40, and a protein subcellular localization benchmark, none of which appears in the definition of D2 or in the constraint. No fitted parameter is fed back into a prediction equation: the epsilon bounds are chosen by validation-set hyperparameter search (Appendix B), and the reported SWGG levels in Appendix D are consequences of optimizing the Lagrangian, not predictions of accuracy. Section 3 explicitly labels the small-D2-good-slice premise as a hypothesis ('we hypothesize that a 'good' slice ... is one for which the SWGG dissimilarity ... is as small as possible'), so the paper does not present it as derived from first principles. The minor self-citations to the authors' prior work on expected sliced transport plans [59] and the softsort-based Espformer [93] are background or implementation references and are not load-bearing for the empirical comparison; the D2 formula is displayed in the paper and the softsort relaxation is also credited to the external reference [87]. The skeptic's concern that only soft-sort SWGG values are reported during training, while inference uses hard argsort permutations, is a correctness or validation gap rather than a circularity: the accuracy comparison does not assume hard constraint satisfaction, and the observed gains could in principle be attributed to other training components, but that is an attribution issue, not an equation-level reduction of output to input. Overall, no step in the derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities; its free parameters are standard hyperparameters (epsilon, alpha, tau, M, learning rates), and the axioms are the SWGG upper-bound property, the informativeness hypothesis, and the faithfulness of the primal-dual and softsort optimizations.

free parameters (6)
  • constraint upper bound epsilon = image: 21; point cloud: 7; subcellular: 5 or 10 per model/slice count
    Chooses the maximum permitted SWGG dissimilarity per slice; tuned on validation grids (Table 2) and directly controls the feasible region of the constrained problem.
  • slack regularization coefficient alpha = image: 0.1; point cloud: 1; subcellular: 0.1 or 1
    Penalizes slack variable magnitude in objective (11a); selected by validation.
  • softsort temperature tau = image: 0.01; point cloud: 0.001; subcellular: 0.001 or 0.01
    Controls how close the differentiable permutation approximation (23) is to a hard argsort; affects gradient quality of the constraint.
  • reference set size M = image: 196; point cloud: 512; subcellular: 100
    Number of trainable reference embeddings in the SWE pool; set to token count for images and point clouds, fixed to 100 for proteins.
  • dual learning rate eta_lambda = 0.001 (all tasks; 0.01 in some grids)
    Step size for the dual variable ascent in Algorithm 1; tuned per task.
  • slack learning rate eta_s = 0.001 (image, point cloud); 0.01 (subcellular)
    Step size for the slack variable descent in Algorithm 1; tuned per task.
assumptions (4)
  • standard math W2(mu,nu) <= D2(mu,nu;theta) for all theta, i.e., SWGG is an upper bound on the true OT cost
    Stated in Section 2.2, imported from Mahey et al. [64]; relies on the quotient-space construction of the lifted plan.
  • domain assumption A slice is informative for downstream learning if its SWGG dissimilarity D2 with the reference set is small
    The paper's central hypothesis, stated in Section 3: 'we hypothesize that a good slice is one for which the SWGG dissimilarity D2 is as small as possible'. Not derived; validated empirically on three tasks.
  • domain assumption The primal-dual stochastic procedure converges to a solution of the relaxed constrained problem
    Algorithm 1 follows the primal-dual constrained learning framework of Chamon/Ribeiro [14,15] and Hounie et al. [37], but no convergence proof is given for the non-convex, stochastic setting.
  • domain assumption The softsort matrix approximates the hard permutation matrix closely enough that gradients of the constraint lead to slices with low true SWGG
    Section 4.1 replaces hard permutation matrices with softsort (Eq. 23) to make the constraint differentiable; the paper does not verify on hard permutations that learned slices satisfy the epsilon bounds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Constrained Sliced Wasserstein Embedding." pith.science (2026). https://pith.science/paper/X2ZPG47U

@misc{pith2026250602203,
  author       = {Pith},
  title        = {Pith review of: Constrained Sliced Wasserstein Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2ZPG47U}},
  note         = {Machine review of arXiv:2506.02203}
}
read the original abstract

Sliced Wasserstein (SW) distances offer an efficient method for comparing high-dimensional probability measures by projecting them onto multiple 1-dimensional probability distributions. However, identifying informative slicing directions has proven challenging, often necessitating a large number of slices to achieve desirable performance and thereby increasing computational complexity. We introduce a constrained learning approach to optimize the slicing directions for SW distances. Specifically, we constrain the 1D transport plans to approximate the optimal plan in the original space, ensuring meaningful slicing directions. By leveraging continuous relaxations of these transport plans, we enable a gradient-based primal-dual approach to train the slicer parameters, alongside the remaining model parameters. We demonstrate how this constrained slicing approach can be applied to pool high-dimensional embeddings into fixed-length permutation-invariant representations. Numerical results on foundation models trained on images, point clouds, and protein sequences showcase the efficacy of the proposed constrained learning approach in learning more informative slicing directions. Our implementation code can be found at https://github.com/Stranja572/constrainedswe.

Figures

Figures reproduced from arXiv: 2506.02203 by the authors.

Figure 1
Figure 1. (Left) Example of SWGG values for a pair of distribu [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Classification accuracy of con￾strained vs. unconstrained SWE on Tiny ImageNet. Means and standard devia￾tions are reported based on three runs. We first consider the task of image classification using DeiT-Tiny [101], a Vision Transformer (ViT) trained and fine-tuned on ImageNet1k [23] with 12 transformer layers and a classifier layer. We freeze the backbone transformer layers and train a classifier on the pooled e… view at source ↗
Figure 3
Figure 3. Test accuracies of PCT on Mod￾elNet40 using unconstrained/constrained SWE and GAP. Means and standard de￾viations are reported based on three runs. To evaluate the effectiveness of constrained SWE in point cloud classification, we conduct experiments us￾ing Point Cloud Transformers (PCT) [33] on the Mod￾elNet40 dataset [109], comprising 3D CAD models from 40 object categories. For each model, we sample 512 points to… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Test accuracy of the proposed method as compared to unconstrained SWE and CLS token [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: (a) Top-5 classification accuracy of constrained vs. unconstrained SWE on Tiny ImageNet, and [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: The evolution of SWGG levels, slack variables, and dual variables in the subcellular localization [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

121 extracted references · 64 canonical work pages

  1. [24]

    Deshpande, Y.-T

    I. Deshpande, Y.-T. Hu, R. Sun, A. Pyrros, N. Siddiqui, S. Koyejo, Z. Zhao, D. Forsyth, and A. G. Schwing. Max-sliced Wasserstein distance and its use for GANs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10648–10656, 2019

  2. [79]

    Nguyen, N

    K. Nguyen, N. Ho, T. Pham, and H. Bui. Distributional sliced-Wasserstein and applica- tions to generative modeling. InInternational Conference on Learning Representations,

  3. [78]

    Nguyen and N

    K. Nguyen and N. Ho. Energy-based sliced wasserstein distance. InThirty- seventh Conference on Neural Information Processing Systems, 2023. URLhttps: //openreview.net/forum?id=umvV3yvo4N

  4. [80]

    Nguyen, T

    K. Nguyen, T. Ren, and N. Ho. Markovian sliced wasserstein distances: Beyond independent projections.Advances in Neural Information Processing Systems, 36: 39812–39841, 2023

  5. [1]

    Adamczyk, A

    J. Adamczyk, A. Arriojas, S. Tiomkin, and R. V. Kulkarni. Utilizing prior solutions for reward shaping and composition in entropy-regularized reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 6658–6665, 2023

  6. [2]

    Adrai, G

    T. Adrai, G. Ohayon, M. Elad, and T. Michaeli. Deep optimal transport: A practical algorithm for photo-realistic image restoration.Advances in Neural Information Processing Systems, 36:61777–61791, 2023

  7. [3]

    J. J. Almagro Armenteros, C. K. Sønderby, S. K. Sønderby, H. Nielsen, and O. Winther. DeepLoc: prediction of protein subcellular localization using deep learning.Bioinfor- matics, 33(21):3387–3395, 2017

  8. [4]

    Alvarez Melis and N

    D. Alvarez Melis and N. Fusi. Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems, 33, 2020

Show all 121 references
  1. [5]

    Amir and N

    T. Amir and N. Dym. Fourier sliced-wasserstein embedding for multisets and measures. arXiv preprint arXiv:2504.02544, 2025

  2. [6]

    Arjovsky, S

    M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein generative adversarial networks. InInternational conference on machine learning, pages 214–223. PMLR, 2017

  3. [7]

    Q. Bai, A. S. Bedi, M. Agarwal, A. Koppel, and V. Aggarwal. Achieving zero constraint violationforconstrainedreinforcementlearningviaprimal-dualapproach. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3682–3689, 2022

  4. [8]

    Bonneel and J

    N. Bonneel and J. Digne. A survey of optimal transport for computer graphics and computer vision. InComputer Graphics Forum, volume 42, pages 439–460. Wiley Online Library, 2023

  5. [9]

    Bonneel, M

    N. Bonneel, M. Van De Panne, S. Paris, and W. Heidrich. Displacement interpola- tion using lagrangian mass transport. InProceedings of the 2011 SIGGRAPH Asia conference, pages 1–12, 2011

  6. [10]

    Bonneel, J

    N. Bonneel, J. Rabin, G. Peyré, and H. Pfister. Sliced and radon wasserstein barycenters of measures.Journal of Mathematical Imaging and Vision, 51:22–45, 2015

  7. [11]

    S. P. Boyd and L. Vandenberghe.Convex optimization. Cambridge university press, 2004

  8. [12]

    Calvo-Fullana, S

    M. Calvo-Fullana, S. Paternain, L. F. Chamon, and A. Ribeiro. State augmented constrained reinforcement learning: Overcoming the limitations of learning with rewards. IEEE Transactions on Automatic Control, 69(7):4275–4290, 2023

  9. [13]

    Chakraborty, D

    S. Chakraborty, D. Paul, and S. Das. Hierarchical clustering with optimal transport. Statistics & Probability Letters, 163:108781, 2020

  10. [14]

    Chamon and A

    L. Chamon and A. Ribeiro. Probably approximately correct constrained learning. Advances in Neural Information Processing Systems, 33:16722–16735, 2020. 18 Constrained Sliced W asserstein Embedding

  11. [15]

    L. F. Chamon, S. Paternain, M. Calvo-Fullana, and A. Ribeiro. Constrained learning with non-convex losses.IEEE Transactions on Information Theory, 69(3):1739–1760, 2022

  12. [16]

    G. Chen, W. Yao, X. Song, X. Li, Y. Rao, and K. Zhang. PLOT: Prompt learning with optimal transport for vision-language models. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum? id=zqwryBoXYnh

  13. [17]

    X. Chen, Y. Yang, and Y. Li. Augmented sliced wasserstein distances.arXiv preprint arXiv:2006.08812, 2020

  14. [18]

    G. Chou, D. Berenson, and N. Ozay. Learning constraints from demonstrations. In International Workshop on the Algorithmic Foundations of Robotics, pages 228–245. Springer, 2018

  15. [19]

    Courty, R

    N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy. Optimal transport for domain adaptation.IEEE transactions on pattern analysis and machine intelligence, 39(9): 1853–1865, 2016

  16. [20]

    M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems, 26, 2013

  17. [21]

    Dadashi, L

    R. Dadashi, L. Hussenot, M. Geist, and O. Pietquin. Primal wasserstein imitation learning. InICLR 2021-Ninth International Conference on Learning Representations, 2021

  18. [22]

    Dai and U

    B. Dai and U. Seljak. Sliced iterative normalizing flows. InICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models, 2021. URLhttps://openreview.net/forum?id=VmwEpdsvHZ9

  19. [23]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255. IEEE, 2009

  20. [25]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

  21. [26]

    Eisen, C

    M. Eisen, C. Zhang, L. F. Chamon, D. D. Lee, and A. Ribeiro. Learning optimal resource allocations in wireless systems.IEEE Transactions on Signal Processing, 67 (10):2775–2790, 2019. 19 NaderiAlizadeh, Salehi, Liu, and Kolouri

  22. [27]

    Elenter, N

    J. Elenter, N. NaderiAlizadeh, and A. Ribeiro. A Lagrangian duality approach to active learning.Advances in Neural Information Processing Systems, 35:37575–37589, 2022

  23. [28]

    Elenter, N

    J. Elenter, N. NaderiAlizadeh, T. Javidi, and A. Ribeiro. Primal dual continual learning: Balancing stability and plasticity through adaptive memory allocation.arXiv preprint arXiv:2310.00154, 2023

  24. [29]

    Fioretto, P

    F. Fioretto, P. Van Hentenryck, T. W. Mak, C. Tran, F. Baldo, and M. Lombardi. Lagrangian duality for constrained deep learning. InMachine learning and knowledge discovery in databases. applied data science and demo track: European conference, ECML pKDD 2020, Ghent, Belgium, S...

  25. [30]

    Gallego-Posada, J

    J. Gallego-Posada, J. Ramirez, A. Erraqabi, Y. Bengio, and S. Lacoste-Julien. Con- trolled sparsity via constrained optimization or: How I learned to stop tuning penalties and love constraints.Advances in Neural Information Processing Systems, 35:1253–1266, 2022

  26. [31]

    Gangavarapu

    A. Gangavarapu. Enhancing guardrails for safe and secure healthcare ai.arXiv preprint arXiv:2409.17190, 2024

  27. [32]

    F. Gao, X. Wang, Y. Fan, Z. Gao, and R. Zhao. Constraints driven safe reinforcement learning for autonomous driving decision-making.IEEE Access, 2024

  28. [33]

    Guo, J.-X

    M.-H. Guo, J.-X. Cai, Z.-N. Liu, T.-J. Mu, R. R. Martin, and S.-M. Hu. Pct: Point cloud transformer.Computational visual media, 7:187–199, 2021

  29. [34]

    J. B. Hakim, J. L. Painter, D. Ramcharran, V. Kara, G. Powell, P. Sobczak, C. Sato, A. Bate, and A. Beam. The need for guardrails with large language models in medical safety-critical settings: An artificial intelligence application in the pharmacovigilance ecosystem.arXiv pre...

  30. [35]

    Haviv, R

    D. Haviv, R. Z. Kunes, T. Dougherty, C. Burdziak, T. Nawy, A. Gilbert, and D. Pe’er. Wasserstein wormhole: Scalable optimal transport distance with transformer. InForty- first International Conference on Machine Learning, 2024. URLhttps://openreview. net/forum?id=Su0qe33cWA

  31. [36]

    K. Hong, Y. Li, and A. Tewari. A primal-dual-critic algorithm for offline constrained reinforcement learning. InInternational Conference on Artificial Intelligence and Statistics, pages 280–288. PMLR, 2024

  32. [37]

    Hounie, A

    I. Hounie, A. Ribeiro, and L. F. Chamon. Resilient constrained learning.Advances in Neural Information Processing Systems, 36:71767–71798, 2023

  33. [38]

    Huynh, H

    V. Huynh, H. Zhao, and D. Phung. Otlda: A geometry-aware optimal transport approach for topic modeling.Advances in Neural Information Processing Systems, 33: 18573–18582, 2020. 20 Constrained Sliced W asserstein Embedding

  34. [39]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y. Bengio and Y. LeCun, editors,3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. URLhttp://arxiv.org/abs/1412.6980

  35. [40]

    Kolluri, V

    J. Kolluri, V. K. Kotte, M. Phridviraj, and S. Razia. Reducing overfitting problem in machine learning using novel l1/4 regularization method. In2020 4th international conference on trends in electronics and informatics (ICOEI)(48184), pages 934–938. IEEE, 2020

  36. [41]

    Kolouri, S

    S. Kolouri, S. R. Park, and G. K. Rohde. The radon cumulative distribution transform and its application to image classification.IEEE transactions on image processing, 25 (2):920–934, 2015

  37. [42]

    Kolouri, A

    S. Kolouri, A. B. Tosun, J. A. Ozolek, and G. K. Rohde. A continuous linear optimal transport approach for pattern analysis in image datasets.Pattern recognition, 51: 453–462, 2016

  38. [43]

    Kolouri, Y

    S. Kolouri, Y. Zou, and G. K. Rohde. Sliced wasserstein kernels for probability distributions. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5258–5267, 2016

  39. [44]

    Kolouri, G

    S. Kolouri, G. K. Rohde, and H. Hoffmann. Sliced wasserstein distance for learning gaussian mixture models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3427–3436, 2018

  40. [45]

    Kolouri, K

    S. Kolouri, K. Nadjahi, U. Simsekli, R. Badeau, and G. Rohde. Generalized sliced wasserstein distances.Advances in neural information processing systems, 32, 2019

  41. [46]

    Kolouri, N

    S. Kolouri, N. NaderiAlizadeh, G. K. Rohde, and H. Hoffmann. Wasserstein embedding for graph learning. InInternational Conference on Learning Representations, 2021

  42. [47]

    Korotin, V

    A. Korotin, V. Egiazarian, A. Asadulaev, A. Safin, and E. Burnaev. Wasserstein-2 generative networks.arXiv preprint arXiv:1909.13082, 2019

  43. [48]

    Kothapalli, A

    A. Kothapalli, A. Shahbazi, X. Liu, R. Sheng, and S. Kolouri. Equivariant vs. invariant layers: A comparison of backbone and pooling for point cloud classification. In ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling, 2024

  44. [49]

    Kotsilieris, I

    T. Kotsilieris, I. Anagnostopoulos, and I. E. Livieris. Regularization techniques for machine learning and their applications, 2022

  45. [50]

    Laclau, I

    C. Laclau, I. Redko, B. Matei, Y. Bennani, and V. Brault. Co-clustering through optimal transport. InInternational conference on machine learning, pages 1955–1964. PMLR, 2017

  46. [51]

    J. Lee, Y. Lee, J. Kim, A. R. Kosiorek, S. Choi, and Y. W. Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. InInternational Conference on Machine Learning, pages 3744–3753. PMLR, 2019. 21 NaderiAlizadeh, Salehi, Liu, and Kolouri

  47. [52]

    Lefevre, A

    S. Lefevre, A. Carvalho, and F. Borrelli. A learning-based framework for velocity control in autonomous driving.IEEE Transactions on Automation Science and Engineering, 13(1):32–42, 2015

  48. [53]

    S. Levine. Reinforcement learning and control as probabilistic inference: Tutorial and review.arXiv preprint arXiv:1805.00909, 2018

  49. [54]

    F.-Z. Li, A. P. Amini, Y. Yue, K. K. Yang, and A. X. Lu. Feature reuse and scaling: Understanding transfer learning with protein language models. InForty-first Interna- tional Conference on Machine Learning, 2024. URLhttps://openreview.net/forum? id=wdTiuvd0fR

  50. [55]

    X. Li, J. Chen, Y. Chai, and H. Xiong. Gilot: interpreting generative language models via optimal transport. InForty-first International Conference on Machine Learning, 2024

  51. [56]

    Z. Lin, H. Akin, R. Rao, B. Hie, Z. Zhu, W. Lu, N. Smetanin, R. Verkuil, O. Kabeli, Y. Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

  52. [57]

    P. Liu, D. Tateo, H. B. Ammar, and J. Peters. Robot reinforcement learning on the constraint manifold. InConference on Robot Learning, pages 1357–1366. PMLR, 2022

  53. [58]

    X. Liu, Y. Bai, R. D. Martín, K. Shi, A. Shahbazi, B. A. Landman, C. Chang, and S. Kolouri. Linear spherical sliced optimal transport: A fast metric for comparing spherical data.arXiv preprint arXiv:2411.06055, 2024

  54. [59]

    X. Liu, R. D. Martin, Y. Bai, A. Shahbazi, M. Thorpe, A. Aldroubi, and S. Kolouri. Ex- pected sliced transport plans. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=P7O1Vt1BdU

  55. [60]

    Y. Liu, Z. Zhou, and B. Sun. Cot: Unsupervised domain adaptation with clustering and optimal transport. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19998–20007, 2023

  56. [61]

    Loshchilov and F

    I. Loshchilov and F. Hutter. SGDR: Stochastic gradient descent with warm restarts. InInternational Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=Skq89Scxx

  57. [62]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019. URLhttps://openreview.net/forum? id=Bkg6RiCqY7

  58. [63]

    Y. Lu, X. Liu, A. Soltoggio, and S. Kolouri. Slosh: Set locality sensitive hashing via sliced-wasserstein embeddings. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2566–2576, 2024

  59. [64]

    Mahey, L

    G. Mahey, L. Chapel, G. Gasso, C. Bonet, and N. Courty. Fast optimal transport through sliced generalized wasserstein geodesics. InThirty-seventh Conference on 22 Constrained Sliced W asserstein Embedding Neural Information Processing Systems, 2023. URLhttps://openreview.net/f...

  60. [65]

    Marco, D

    A. Marco, D. Baumann, M. Khadiv, P. Hennig, L. Righetti, and S. Trimpe. Robot learning with crash constraints.IEEE Robotics and Automation Letters, 6(2):1439–1446, 2021

  61. [66]

    Massiani, A

    P.-F. Massiani, A. von Rohr, L. Haverbeck, and S. Trimpe. Viability of future actions: Robust reinforcement learning via entropy regularization. InSeventeenth European Workshop on Reinforcement Learning, 2024. URLhttps://openreview.net/forum? id=zP9hpDEzPq

  62. [67]

    C. Meng, J. Yu, J. Zhang, P. Ma, and W. Zhong. Sufficient dimension reduction for classification using principal optimal transport direction.Advances in neural information processing systems, 33:4015–4028, 2020

  63. [68]

    Mialon, D

    G. Mialon, D. Chen, A. d’Aspremont, and J. Mairal. A trainable optimal transport embedding for feature aggregation and its relationship to attention. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum? id=ZK6vTvb84s

  64. [69]

    Mialon, D

    G. Mialon, D. Chen, A. d’Aspremont, and J. Mairal. A trainable optimal transport embedding for feature aggregation and its relationship to attention. InICLR 2021-The Ninth International Conference on Learning Representations, 2021

  65. [70]

    E. F. Montesuma, F. M. N. Mboula, and A. Souloumiac. Recent advances in optimal transport for machine learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  66. [71]

    Moradi, R

    R. Moradi, R. Berangi, and B. Minaei. A survey of regularization strategies for deep models.Artificial Intelligence Review, 53(6):3947–3986, 2020

  67. [72]

    Moskovitz, M

    T. Moskovitz, M. Arbel, F. Huszar, and A. Gretton. Efficient wasserstein natural gradients for reinforcement learning. InInternational Conference on Learning Repre- sentations, 2021

  68. [73]

    NaderiAlizadeh and R

    N. NaderiAlizadeh and R. Singh. Aggregating residue-level protein language model embeddings with optimal transport.Bioinformatics Advances, 5(1):vbaf060, 2025

  69. [74]

    NaderiAlizadeh, J

    N. NaderiAlizadeh, J. F. Comer, R. W. Andrews, H. Hoffmann, and S. Kolouri. Pooling by sliced-Wasserstein embedding. InThirty-Fifth Conference on Neural Information Processing Systems, 2021. URLhttps://openreview.net/forum?id=1z2T01DKEaE

  70. [75]

    NaderiAlizadeh, M

    N. NaderiAlizadeh, M. Eisen, and A. Ribeiro. Learning resilient radio resource man- agement policies with graph neural networks.IEEE Transactions on Signal Processing, 71:995–1009, 2023

  71. [76]

    Nadjahi, A

    K. Nadjahi, A. Durmus, L. Chizat, S. Kolouri, S. Shahrampour, and U. Simsekli. Statistical and topological properties of sliced probability divergences.Advances in Neural Information Processing Systems, 33:20802–20812, 2020. 23 NaderiAlizadeh, Salehi, Liu, and Kolouri

  72. [77]

    Nadjahi, A

    K. Nadjahi, A. Durmus, P. Jacob, R. Badeau, and U. Simsekli. Fast approximation of the sliced-wasserstein distance using concentration of random projections. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. W. Vaughan, editors,Advances in Neural Information Processing Systems, ...

  73. [81]

    Nguyen, T

    K. Nguyen, T. Ren, H. Nguyen, L. Rout, T. M. Nguyen, and N. Ho. Hierarchical sliced wasserstein distance. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum?id=CUOaVn6mYEj

  74. [82]

    Nikolentzos, P

    G. Nikolentzos, P. Meladianos, and M. Vazirgiannis. Matching node embeddings for graph similarity. InProceedings of the AAAI conference on Artificial Intelligence, volume 31, 2017

  75. [83]

    O’Donoghue, R

    B. O’Donoghue, R. Munos, K. Kavukcuoglu, and V. Mnih. Combining policy gradient and q-learning. InInternational Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=B1kJ6H9ex

  76. [84]

    G. Oh, B. Sim, H. Chung, L. Sunwoo, and J. C. Ye. Unpaired deep learning for accelerated mri using optimal transport driven cyclegan.IEEE Transactions on Computational Imaging, 6:1285–1296, 2020

  77. [85]

    Paty and M

    F.-P. Paty and M. Cuturi. Subspace robust wasserstein distances. InInternational conference on machine learning, pages 5072–5081. PMLR, 2019

  78. [86]

    Peyré, M

    G. Peyré, M. Cuturi, et al. Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019

  79. [87]

    Prillo and J

    S. Prillo and J. Eisenschlos. Softsort: A continuous relaxation for the argsort operator. InInternational Conference on Machine Learning, pages 7793–7802. PMLR, 2020

  80. [88]

    Rabin, G

    J. Rabin, G. Peyré, J. Delon, and M. Bernot. Wasserstein barycenter and its application to texture mixing. In A. M. Bruckstein, B. M. ter Haar Romeny, A. M. Bronstein, and M. M. Bronstein, editors,Scale Space and Variational Methods in Computer Vision, pages 435–446, Berlin, H...

  81. [89]

    Ramirez, I

    J. Ramirez, I. Hounie, J. Elenter, J. Gallego-Posada, M. Hashemizadeh, A. Ribeiro, and S. Lacoste-Julien. Feasible learning. InThe 28th International Conference on Artificial Intelligence and Statistics, 2025. URLhttps://openreview.net/forum?id= Y1BEQEELxI

  82. [90]

    L. Rout, A. Korotin, and E. Burnaev. Generative modeling with optimal transport maps.arXiv preprint arXiv:2110.02999, 2021

  83. [91]

    Salman and X

    S. Salman and X. Liu. Overfitting mechanism and avoidance in deep neural networks. arXiv preprint arXiv:1901.06566, 2019

  84. [92]

    C. F. G. D. Santos and J. P. Papa. Avoiding overfitting: A survey on regularization methods for convolutional neural networks.ACM Computing Surveys (Csur), 54(10s): 1–25, 2022

  85. [93]

    Shahbazi, E

    A. Shahbazi, E. Akbari, D. Salehi, X. Liu, N. NaderiAlizadeh, and S. Kolouri. Espformer: Doubly-stochastic attention with expected sliced transport plans.arXiv preprint arXiv:2502.07962, 2025

  86. [94]

    Y. Shi, L. Lian, Y. Shi, Z. Wang, Y. Zhou, L. Fu, L. Bai, J. Zhang, and W. Zhang. Machine learning for large-scale optimization in 6g wireless networks.IEEE Communi- cations Surveys & Tutorials, 25(4):2088–2132, 2023

  87. [95]

    Shifat-E-Rabbi, X

    M. Shifat-E-Rabbi, X. Yin, A. H. M. Rubaiyat, S. Li, S. Kolouri, A. Aldroubi, J. M. Nichols, and G. K. Rohde. Radon cumulative distribution transform subspace modeling for image classification.Journal of Mathematical Imaging and Vision, 63:1185–1203, 2021

  88. [96]

    Tiny imagenet

    Stanford University. Tiny imagenet. http://cs231n.stanford.edu/ tiny-imagenet-200.zip

  89. [97]

    Stärk, C

    H. Stärk, C. Dallago, M. Heinzinger, and B. Rost. Light attention predicts protein location from the language of life.Bioinformatics Advances, 1(1):vbab035, 11 2021. ISSN 2635-0041. doi: 10.1093/bioadv/vbab035

  90. [98]

    B. E. Suzek, Y. Wang, H. Huang, P. B. McGarvey, C. H. Wu, and U. Consortium. Uniref clusters: a comprehensive and scalable alternative for improving sequence similarity searches.Bioinformatics, 31(6):926–932, 2015

  91. [99]

    Tian and Y

    Y. Tian and Y. Zhang. A comprehensive survey on regularization strategies in machine learning.Information Fusion, 80:146–166, 2022

  92. [100]

    Togninalli, E

    M. Togninalli, E. Ghisu, F. Llinares-López, B. Rieck, and K. Borgwardt. Wasserstein weisfeiler-lehman graph kernels.Advances in Neural Information Processing Systems, 32:6439–6449, 2019

  93. [101]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou. Training data-efficient image transformers & distillation through attention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021. 25 NaderiAlizadeh, Salehi, Liu, and Kolouri

  94. [102]

    H. Tran, Y. Bai, A. Shahbazi, J. R. Hershey, and S. Kolouri. Understanding learn- ing with sliced-wasserstein requires rethinking informative slices.arXiv preprint arXiv:2411.10651, 2024

  95. [103]

    Van Assel, C

    H. Van Assel, C. Vincent-Cuaz, N. Courty, R. Flamary, P. Frossard, and T. Vayer. Distributional reduction: Unifying dimensionality reduction and clustering with gromov- wasserstein.arXiv preprint arXiv:2402.02239, 2024

  96. [104]

    Villani et al.Optimal transport: old and new, volume 338

    C. Villani et al.Optimal transport: old and new, volume 338. Springer, 2008

  97. [105]

    J. Wang, B. Lei, L. Ding, X. Xu, X. Gu, and M. Zhang. Autoencoder-based conditional optimal transport generative adversarial network for medical image generation.Visual Informatics, 8(1):15–25, 2024

  98. [106]

    L. Wang, X. Li, H. Zhang, J. Wang, D. Jiang, Z. Xue, and Y. Wang. A comprehensive review of protein language models.arXiv preprint arXiv:2502.06881, 2025

  99. [107]

    W. Wang, D. Slepčev, S. Basu, J. A. Ozolek, and G. K. Rohde. A linear optimal transportation framework for quantifying and visualizing variations in sets of images. International journal of computer vision, 101:254–269, 2013

  100. [108]

    Weissenow and B

    K. Weissenow and B. Rost. Are protein language models the new universal key? Current Opinion in Structural Biology, 91:102997, 2025

  101. [109]

    Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015

  102. [110]

    H. Xiao, M. Herman, J. Wagner, S. Ziesche, J. Etesami, and T. H. Linh. Wasserstein adversarial imitation learning.arXiv preprint arXiv:1906.08113, 2019

  103. [111]

    Y. Xiao, W. Zhao, J. Zhang, Y. Jin, H. Zhang, Z. Ren, R. Sun, H. Wang, G. Wan, P. Lu, et al. Protein large language models: A comprehensive survey.arXiv preprint arXiv:2502.17504, 2025

  104. [112]

    K.D.Yang, K.Damodaran, S.Venkatachalapathy, A.C.Soylemezoglu, G.Shivashankar, and C. Uhler. Predicting cell lineages using autoencoders and optimal transport.PLoS computational biology, 16(4):e1007828, 2020

  105. [113]

    Y. Yang, Q. Jin, R. Leaman, X. Liu, G. Xiong, M. Sarfo-Gyamfi, C. Gong, S. Ferrière- Steinert, W. J. Wilbur, X. Li, et al. Ensuring safety and trust: Analyzing the risks of large language models in medicine.arXiv preprint arXiv:2411.14487, 2024

  106. [114]

    Zafar, M

    A. Zafar, M. Aamir, N. Mohd Nawi, A. Arshad, S. Riaz, A. Alruban, A. K. Dutta, and S. Almotairi. A comparison of pooling methods for convolutional neural networks. Applied Sciences, 12(17):8643, 2022

  107. [115]

    Zaheer, S

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdinov, and A. J. Smola. Deep sets.Advances in neural information processing systems, 30, 2017. 26 Constrained Sliced W asserstein Embedding

  108. [116]

    Zhang, P

    J. Zhang, P. Ma, W. Zhong, and C. Meng. Projection-based techniques for high- dimensional optimal transport problems.Wiley Interdisciplinary Reviews: Computa- tional Statistics, 15(2):e1587, 2023

  109. [117]

    Zhang, C

    R. Zhang, C. Chen, C. Li, and L. Carin. Policy optimization as Wasserstein gradient flows. In J. Dy and A. Krause, editors,Proceedings of the 35th International Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pages 5737–5746. PMLR, 10–15 J...

  110. [118]

    Zhang, C

    R. Zhang, C. Chen, Z. Gan, Z. Wen, W. Wang, and L. Carin. Nested-wasserstein self-imitation learning for sequence generation. InInternational Conference on Artificial Intelligence and Statistics, pages 422–433. PMLR, 2020

  111. [119]

    Zhang, X

    Y. Zhang, X. Liang, D. Li, S. S. Ge, B. Gao, H. Chen, and T. H. Lee. Adaptive safe reinforcement learning with full-state constraints and constrained adaptation for autonomous vehicles.IEEE Transactions on Cybernetics, 54(3):1907–1920, 2023

  112. [120]

    R. Zhao, X. Sun, and V. Tresp. Maximum entropy-regularized multi-goal reinforcement learning. InInternational Conference on Machine Learning, pages 7553–7562. PMLR, 2019. 27

  113. [2021]

    URLhttps://openreview.net/forum?id=QYjO70ACDK

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.