acceptodds
Withdrawn from ICLR 2027

GENERATIVE MODEL VIA QUANTILE ASSIGNMENT

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

Learning complex high-dimensional distributions is a challenging problem, and modern generative models address it through a variety of architectures and training objectives, often involving auxiliary learned components such as encoders, discriminators, or critics. We introduce neuroSQL, a generative modeling approach whose only learned component is a single neural generator. neuroSQL fixes a lattice of multivariate quantile points in latent space and trains the generator by alternating between a regression step and an assignment step that matches data points to latent quantiles. Under squared Euclidean cost, exact assignment yields the optimal transport coupling between the empirical data measure and the generator's pushforward of the quantile lattice, so the training objective is exactly their squared Wasserstein-2 distance. This explicit objective makes the optimization amenable to analysis: under suitable smoothness and step-size conditions, we prove that the alternating procedure decreases the objective monotonically and reaches a stationary point of it at rate — a guarantee unavailable to adversarial training, which has no single potential function, or to variational training, which is monotone only on a bound. The resulting algorithm is lightweight, fast, and deterministic given the lattice. Controlled two-dimensional experiments reach held-out sliced Wasserstein distances comparable to an independent real–real reference. A million-observation simulation confirms that block assignment keeps cost-matrix memory bounded and solver cost linear in . When trained on all 60,000 MNIST training images using block assignment, neuroSQL recovers all ten classes with no exact training-image copies, and is competitive with matched-backbone VAE and GLO baselines at lower training cost. Our guarantees are finite-sample optimization guarantees rather than population-consistency results, and within-class coverage remains the method's clearest limitation.

Primary area:
generative models
Keywords:
Generative models, Optimal transport, Wasserstein distance

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_3VhoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an optimal-transport-based generative model as an alternative to classical one-step generative models, such as VAEs. Therefore, it proposes to discretize the latent space via a discretized quantile functions to fit an OT assignment between latent quantile and data points. New samples are generated via sampling new point form the latent distribution.

Critical strengths

The idea of using quantiles to discretize the latent space with respect to the latent space has not been explored that much.

In comparison to other OT-inspired classical one-step generative models, such as Wasserstein autoencoders and Wasserstein GANs, the connection of the proposed model to optimal transport seems closer in spirit.

Critical weaknesses

Discretization of the quantile assignment is hard to scale to higher-dimensions due to the curse of dimensionality.

While the idea is interesting, the model is most certainly not competitive with current one-step model. While I do not think that performance is the only evaluation metric, the underlying theory is not sufficient to justify interest in the model (in my personal opinion).

The model is not even compared to other OT-inspired methods: MinibatchOT Flow Matching (Tong et al), Neural Optimal Transport (Korotin et al), Wasserstein Autoencoders (Tolstikhin et al), ...

The fundamental definition of a multivariate quantile function is not explained in the main text.

There is a link that points to "Anonymous GitHub" for code, but I could not find anything.

Minor:

  • One should clarify that iterated optimization procedure wrt to theta and pi does not lead to global optimization. Additionally, blockwise optimization leads to further relaxation.
Questions for authors

Given that I am only supposed to ask "questions whose answers could materially affect your assessment," I have no questions.

While the comments above reflect my own assessment, I would like to point out that ICLR offers a rather good LLM-based paper precheck service (the "Paper Assistant Tool," or PAT). Although my main concerns are conceptual, I think the paper could have benefited from careful use of this service.


On a more personal note, I think the proposed model builds on some genuinely interesting ideas! :) Given the remarkable progress in generative modeling over the past years, and the steady refinement of model architectures from VAEs and GANs to flows and diffusion models, competing with established approaches is certainly no easy task. I hope the authors find my comments helpful in further developing their ideas.

Reviewer_o9mdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a generative model named neuroSQL, whose only learning part is the generator. neuroSQL works in a latent space, fixes a lattice of multivariate quantiles of the prior, and represents the unknown latent realization as a permutation of the lattice. Theoretically, the authors show that the training objective is essentially the squared W2 distance between quantile and data empirical distributions. They also show the monotone decreasing of the objective under Lipschitz assumptions. When the assumptions fail to hold, a Clarke gradient is used. Experiments use MNIST, toy data and CelebA and compare with GLO and VAE.

Critical strengths
  1. This paper is based on a simple but interesting idea. The idea could provide some insights for a general audience of ICLR.
  2. The implementation requires alternating minimization and is shown to converge given assumptions. This could make it a potential approach for solving the instability in other generator training processes.
  3. The paper is well-written. It is easy to follow the logic and idea. Figure 1 is also helpful for understanding the algorithm.
Critical weaknesses
  1. The core idea is based on multivariate quantiles. But it seems that way the quantiles are constructed is directly taken from (Hallin et al., 2021). So the main contribution is restricted to the algorithm design and convergence analysis.
  2. In the experiments, the authors only compare neuroSQL with some very simple baselines like VAE. The main test focuses on simple datasets like MNIST. This makes it uncertain the practical usage of the method.
  3. The authors may consider using the same strategy within more advance architectures to make the quality of the generated images better.
Questions for authors

See "weaknesses".

Reviewer_DcZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NEUROSQL trains a generator by alternating regression and bijective assignment to a fixed latent quantile lattice. The paper provides a Wasserstein interpretation and optimization guarantees for exact assignment, and evaluates a scalable implementation on synthetic data, MNIST, and CelebA.

Critical strengths
  1. The formulation is intuitive and admits a clear analysis under exact assignment.

  2. The MNIST and CelebA results are promising and provide meaningful real-image evidence.

Critical weaknesses
  1. Earlier work, particularly Chen et al. (ICML 2019), already combines optimal transport and generator regression. The paper should explain precisely what its quantile-based formulation adds. Moreover, the Wasserstein identity and descent analysis apply to arbitrary fixed, equally weighted latent points, so they do not establish an advantage specific to quantiles.

  2. The image pipeline combines quantile assignment with descriptor matching, PCA initialization, compound losses, jitter, and EMA. Without controlled ablations, the improvements cannot be attributed to the proposed core mechanism. The theory analyzes an exact, same-cost procedure and therefore does not directly explain the gains of this practical implementation.

  3. Efficient generative modeling is important, but the specific bottleneck addressed by this design remains insufficiently demonstrated. Removing auxiliary networks introduces assignment costs, making end-to-end evaluation essential. Comparisons with the selected VAE/GLO implementations alone do not establish a meaningful advantage in sample efficiency, coverage, or generation quality relative to computational cost. Closer baselines, including IMLE and direct OT-based training, would help assess this significance.

Questions for authors
  1. What is the essential advance over earlier transport–regression methods, and why should quantiles outperform fixed i.i.d. latent points?
  2. Which components account for the gains when initialization, losses, and optimization settings are controlled?

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

Related papers

Discussion (0)

Sign in to comment.