acceptodds
Withdrawn from ICLR 2027

Neural Adaptive Optimal Transport: Duality, Brenier Potentials and Applications

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Classical optimal transport is a cornerstone of modern machine learning, yet its rigid requirement for full-mass or fixed-mass conservation severely limits its robustness against outliers and distributional shifts. While recent discrete elastic optimal transport enables adaptive mass transport via inequality constraints and mixed-sign costs, it lacks a continuous theoretical foundation, leaving critical questions on duality, Brenier potentials, and high-dimensional solvability unanswered. To bridge this gap, we propose Neural Adaptive Optimal Transport (NAOT), a novel framework that unifies discrete elastic optimal transport with continuous optimal transport theory. Building upon Brenier's seminal theory, we derive the duality theory and Brenier-type theorems, which shed light on designing an efficient and scalable algorithm based on input convex neural networks to find transport map. NAOT outperforms the state-of-the-art methods at predicting single-cell perturbation responses on the benchmark dataset, demonstrating its superiority on adaptive distribution alignment. Our work provides the first continuous framework of adaptive optimal transport, offering both theoretical insights and a scalable algorithm with broad applicability in artificial intelligence.

Primary area:
learning theory
Keywords:
adaptive optimal transport, single-cell perturbation, duality theory, neural optimal transport, input convex neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Hb7uOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces neural adaptive optimal transport, in which transport cost consists of two parts: first component is standard Euclidean cost, and second component is the task dependent learnable projection matrix. The paper specifically generalizes the ideas of recently developed elastic optimal transport for discrete optimal transport to continuous settings by employing ideas from Brenier's and partial optimal transport to derive the optimization objective which is amenable to neural network-based parametrization. Drawing inspiration derived results and previous work on neural optimal transport, the paper parametrizes dual potentials using input convex neural networks.

Critical strengths

The paper is very well written and easy to follow and discusses an interesting variant of optimal transport, which can have wide variety of applications.

Critical weaknesses
  1. Paper seems to be theoretical in nature, but only true theoretical contribution of paper appears to come from theorem 1, where it adopts the ideas employed in derivation of fixed-budget partial optimal transport problem by Caffarelli and McCann [1]. Remainder of the theorems in paper are adoptions of general optimal transport theory for the cost function introduced in equation 4, which severely limits the contributions of paper.
  2. As far experimental results are concerned, all experiments are done on a single-cell perturbation dataset, whereas similar frameworks developed for optimal transport involve applications on multiple datasets [2].

[1] Caffarelli, Luis A., and Robert J. McCann. "Free boundaries in optimal transport and Monge-Ampere obstacle problems." Annals of mathematics (2010): 673-730.

[2] Korotin, Alexander, Daniil Selikhanovych, and Evgeny Burnaev. "Neural optimal transport." arXiv preprint arXiv:2201.12220 (2022).

Questions for authors

For paper to be considered seriously, I would expect the authors to address above mentioned weaknesses in addition to providing satisfactory answers to questions asked below.

  1. Could authors clearly explain where does the adaptivity the proposed in framework appear. I know cost function has learnable Matrix M, but I want to know and understand where that does the adaptivity appear in the proof of theorem 1, which is essentially the core of methodology developed in the paper.
  2. From equation 4, it seems that the projection matrix M is constant, but the discussion in appendix E, suggests otherwise. Please clarify that.
  3. In neural optimal transport literature, it has been observed multiple times that although ICNN have nice theoretical properties they are difficult to initialize and train [2]. How did you deal with that?
  4. In Neural optimal transportation, when transport operator is directly parametrized, it suffers from spurious solutions problem, does that happen in your case [1,2] ? If yes, what do you suggest to resolve that?

[1] Choi, Jaemoo, Jaewoong Choi, and Dohyun Kwon. "Overcoming spurious solutions in semi-dual neural optimal transport: A smoothing approach for learning the optimal transport plan." arXiv preprint arXiv:2502.04583 (2025).

[2] Rout, Litu, Alexander Korotin, and Evgeny Burnaev. "Generative modeling with optimal transport maps." arXiv preprint arXiv:2110.02999 (2021).

Reviewer_mHPROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a new approach which lies in the field of continuous optimal transport (OT) solvers. Their approach is based on the elastic optimal transport formulation (ELOT, Yang et al., 2026) which relaxes the constraints of classical OT problem. Continuous solvers based on elastic formulation were absent in the field and the authors fill this gap. Their solver is based on Input Convex Neural Networks (ICNNs) and is tested in several biological experiments.

Critical strengths

The paper is well-written and easy to follow. The developed method fills a gap in existing continuous OT-based solvers by extending recent elastic OT formulation to the continuous setting.

Critical weaknesses

[W1] The main motivation of extending elastic OT framework to continuous setting was explained (see Introduction and Abstract) by the fact that this OT formulation relaxes the strict constraints in the classical OT problem and, thus, should be more robust to outliers and other imperfections in input data. Well-known extensions of the OT problem offering such kind of robustness include partial and unbalanced OT. The authors claim that elastic (or adaptive) OT offers another kind of flexibility which explains the importance of developing the related continuous solver. Thus, it is natural to expect that the paper will thoroughly analyse the differences and show the advantages of the developed approach in comparison to unbalanced and/or partial OT-based ones. However, actually, the method seems to be tailored to the particular biological tasks since all of the experiments are biological and even the cost function which is considered in most of the theoretical results seems to be motivated by the biological experiments considered further. Thus, I think that the paper should be enhanced by either considering a more broad scope of applications or stating the biological focus from the beginning. Maybe, this should be obvious even from the title - the paper of (Klein et al., 2024) is a good example.

[W2] Hereafter, I detail my concerns regarding the considered experiments taking into account that in the current version of the paper the method is positioned as a general one. The first significant gap is the absence of tests in toy experiments showing that (1) the method allows for learning the actual elastic transport maps and (2) it has differences and advantages over partial and unbalanced OT solvers in terms of robustness to outliers, class imbalance or other data imperfections. I kindly suggest the authors include these experiments in the paper.

[W3] In the biological experiments, the paper compares the proposed NAOT approach with several other OT-based methods, i.e., CellOT (Bunne et al., 2023), CondOT (Bunne et al., 2022) and GENOT (Klein et al., 2024). However, the paper does not sufficiently analyze the particular properties of the designed NAOT solver’s solutions in comparison with unbalanced or partial ones. In section 4.1, the authors claim that their approach leads to “significant performance gain” while from Table 1 it is obvious that the unbalanced OT-based method, GENOT, leads to better performance in Trametinib compound. Similar situation appears in experimental section 4.3, where the NAOT approach is beaten by other methods w.r.t. some metrics on both compounds. Why does this happen? How can the obtained results be explained from a biological point of view? Is there any explanation why relaxation of NAOT approach is meaningful in this setup and why it is better than unbalanced or partial ones?

[W4] The related issue is the limited number of considered baselines in section 4 - two of three competitors considered here correspond to classical OT case. This contradicts with the main idea of the paper where the authors claim to develop a continuous solver with a new type of relaxation. Thus, I believe that more unbalanced and unbalanced solvers should be considered here. There are a lot of papers on this topic published in recent years (Choi et al, 2023; Eyring et al., 2024; Lübeck et al., 2022; Gazdieva et al., 2024; Yang and Uhler, 2018) and I kindly suggest the authors to add some of them into comparison.

[W5] Importantly, in the experiments, the authors consider their NAOT approach conditioned on the biological context, e.g. drug dosages, and the related weakness is an unfair comparison with original CellOT and GENOT approaches which do not use conditioning by default. I kindly ask the authors to correct me, if I am wrong in understanding that they did not include conditioning into implementations of these approaches. (There are no realizations of these methods in the code and no explanations in the paper supporting this idea.) However, if I am correct in my understanding, then the generalization ability of the NAOT approach shown in experiment from section 4.2 can be treated just as a consequence of its conditional nature. Thus, I kindly suggest the authors consider the conditional versions of the baseline methods to ensure fair comparison here.

[W6] One more concern is related to the construction of the approach where the constraints in objective (14) are relaxed using the penalty term (16). It is not obvious whether the solutions of the final objective will be close to the solutions of the initial one. Thus, it is unclear whether the approach indeed allows for learning true adaptive plans. I kindly suggest the authors consider establishing some theoretical guarantee, e.g., deriving bounds on the discrepancy between the solutions of the initial and relaxed problems.

Overall, I see the significant gaps between the motivation of developing the adaptive solver and theoretical and, especially, experimental justifications which are particularly tailored to biological tasks. This raises questions regarding the overall positioning of the paper results, see the details above.

References.

Pei Yang, Yuhang Zhuang, and Qi Tan. Elastic optimal transport: Theory, application, and empirical evaluation. In The Fourteenth International Conference on Learning Representations (ICLR), 2026.

Charlotte Bunne, Andreas Krause, and Marco Cuturi. Supervised training of conditional monge maps. In Neural Information Processing Systems (NeurIPS), volume 35, pp. 6859–6872, 2022.

Charlotte Bunne, Stefan G Stark, Gabriele Gut, Jacobo Sarabia Del Castillo, Mitch Levesque, Kjong-Van Lehmann, Lucas Pelkmans, Andreas Krause, and Gunnar Ratsch. Learning single-cell perturbation responses using neural optimal transport. Nature Methods, 20:1759–1768, 2023.

Dominik Klein, Theo Uscidda, Fabian Theis, and Marco Cuturi. Genot: Entropic (gromov) wasser-stein flow matching with applications to single-cell genomics. Advances in Neural Information Processing Systems, 37:103897–103944, 2024.

J. Choi, J. Choi, and M. Kang. Generative modeling through the semi-dual formulation of unbalanced optimal transport. In Advances in Neural Information Processing Systems, volume 36, 2023.

L. Eyring, D. Klein, T. Uscidda, G. Palla, N. Kilbertus, Z. Akata, and F. Theis. Unbalancedness in neural monge maps improves unpaired domain translation. In The Twelfth International Conference on Learning Representations, 2024.

F. Lübeck, C. Bunne, G. Gut, J. S. del Castillo, L. Pelkmans, and D. Alvarez-Melis. Neural unbalanced optimal transport via cycle-consistent semi-couplings. arXiv preprint arXiv:2209.15621, 2022.

M. Gazdieva, A. Asadulaev, E. Burnaev, A. Korotin. Light unbalanced optimal transport. Advances in Neural Information Processing Systems, 2024

K. D. Yang and C. Uhler. Scalable unbalanced optimal transport using generative adversarial networks. In International Conference on Learning Representations, 2018.

Questions for authors

See Weaknesses section.

Reviewer_h6XrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Neural Adaptive Optimal Transport (NAOT), a continuous formulation of elastic optimal transport that minimizes a mixed-sign transport cost subject to upper bounds on both marginals, allowing the transported mass to be selected rather than fixed in advance. For a quadratic cost augmented with a learned projected-space similarity term, the paper claims duality results and a Brenier-type characterization of transport maps. It then develops an ICNN/PICNN-based min–max solver that obtains conditional transport maps from gradients of learned potentials. Validation focuses on single-cell perturbation prediction using sciPlex3 subsets for givinostat and trametinib. The experiments compare NAOT with CELLOT, CondOT, and GENOT across dosage-conditioned and cell-line-conditioned tasks, including a held-out 100 nM dosage, using MMD, Wasserstein distance, perturbation-signature error, and FID.

Critical strengths

The reported held-out 100 nM results show lower MMD and FID than the evaluated baselines for both drugs, providing a positive empirical signal for conditional distribution prediction.

Critical weaknesses
  1. Equation (5) requires both quadratic lower bounds, in addition to the coupling constraint:
\psi(z)\geq \tfrac12\\|z\\|^2,\qquad \phi(x)+\psi(z)\geq x^\top Qz.$$ For a fixed feasible $\phi$, the pointwise smallest admissible second potential is therefore $$\psi(z)=\max\left\\{\tfrac12\\|z\\|^2,\\ \phi^*(z)\right\\},$$ where $\phi^*$ denotes the paper’s $Q$-dependent conjugate, rather than simply $\phi^*(z)$. However, Appendix C.3 asserts mutual conjugacy in Equations (37)–(38) without accounting for these lower bounds. Appendix C.4 then derives an unconstrained conjugate-based min–max representation before adding the lower bounds back. This does not establish the claimed equivalence. The correct elimination retains the maximum above, whereas the proposed inner objective depends only on $\nabla\psi$. On a compact domain, shifting $\psi$ upward can satisfy its lower bound without changing its gradient, so imposing that bound afterward cannot recover the missing term. Theorems 4–5 therefore do not establish that the neural objective solves the stated adaptive-mass problem. 2. Equations (16)–(17) add a nonnegative constraint-violation penalty with a positive coefficient while maximizing over $\psi$: $$\min_\phi\max_\psi\\; J_{\mathrm{emp}}(\phi,\psi)+\sigma R_c(\phi,\psi).$$ This rewards violations of the lower bound on $\psi$. More concretely, let $\psi_K(z)=\psi(z)-K$. This shift preserves convexity and leaves $\nabla\psi$, and hence $J_{\mathrm{emp}}$, unchanged. The corresponding penalty becomes $$\frac1m\sum_j \left[\tfrac12\\|z_j\\|^2-\psi(z_j)+K\right]_+^2,$$ which diverges as $K\to\infty$. Thus, for $\sigma>0$, the inner maximization is unbounded over the stated convex function class. Algorithm 1 also instructs the $\psi$-network to maximize Equation (17), so the inconsistency appears in both the objective and pseudocode. The exact implemented losses need to be specified and reconciled with the reported results. Correcting a possible sign error would still leave the separate loss of the adaptive-mass constraints in the theoretical reduction. 3. The primal permits transported submarginals $\gamma_X\leq\mu$ and $\gamma_Z\leq\nu$. A map characterization on the transported mass would concern $$T_\\#\gamma_X=\gamma_Z,$$ not automatically a map transporting all of $\mu$ to all of $\nu$ or a global inverse, as stated in Theorem 3. Nevertheless, inference maps every control cell and evaluates an equally weighted predicted population, without explaining how inactive mass is represented. The authors need to clarify the scope of the map theorem and show how the solver recovers transported mass or selection weights. Normalized distribution-fitting results alone do not validate the claimed modeling of proliferation and death.
Questions for authors
  1. Can you derive Equations (11)–(13) from Equation (5) while retaining both quadratic lower bounds, and explain how the resulting maximum term is represented in the gradient-only inner objective? Does Theorem 3 concern transported submarginals or the complete input measures, and under what conditions does its inverse-map statement hold?

  2. Does the implementation maximize +σRc+\sigma R_c in the ψ\psi-update, as Equation (17) and Algorithm 1 specify? Please provide the exact loss for each network, including all penalty signs. If the implementation differs from the manuscript, which objective produced the reported results, and how does it avoid the constant-shift unboundedness above?

  3. What transported mass and source/target selection weights does the implementation recover? A small controlled comparison against an exact solution of Equation (1), reporting transported mass, primal cost, and marginal feasibility, would materially affect my assessment. In particular, at α=0\alpha=0, the cost is nonnegative, so γ=0\gamma=0 is always optimal and is uniquely optimal for disjoint supports. How is this reconciled with Appendix E.3’s attribution of favorable α=0\alpha=0 prediction performance to the adaptive-mass mechanism?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.