acceptodds
Withdrawn from ICLR 2027

Discrete Diffusion with Sample-Efficient Estimators for Conditionals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We study a discrete denoising diffusion framework that integrates a sample-efficient estimator of single-site conditionals with round-robin noising and denoising dynamics for generative modeling over discrete state spaces. Rather than approximating a discrete analog of a score function, our formulation treats single-site conditional probabilities as the fundamental objects that parameterize the reverse diffusion process. We employ a sample-efficient method known as Neural Interaction Screening Estimator (NeurISE) to estimate these conditionals in the diffusion dynamics. Controlled experiments on synthetic Ising models, MNIST, and scientific data sets produced by a D-Wave quantum annealer, synthetic Potts model and one-dimensional quantum systems demonstrate the proposed approach. On the binary data sets, these experiments demonstrate that the proposed approach outperforms popular existing methods including ratio-based approaches, achieving improved performance in total variation, cross-correlations, and kernel density estimation metrics.

Primary area:
generative models
Keywords:
Diffusion models, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_uGTBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper formulates a machinery for training discrete diffusion models for variable σ=[σ1,σ2,…,σq]∈Σq\sigma = [\sigma_1, \sigma_2, \dots, \sigma_q] \in \Sigma^q. Their basic strategy is as follows:

  1. Suppose data distribution as μ0\mu_0,

  2. Corrupt it by a kernel knk_n as μn+1(σ)=∑σ~kn(σ,σ~)μn(σ~)\mu_{n+1}(\sigma) = \sum_{\tilde{\sigma}} k_n(\sigma, \tilde{\sigma}) \mu_n(\tilde{\sigma}) where knk_n is "round-robin" forward noising defined as kn(σ,σ~)={≠0∃u, s.t. [σ1,…,-,…,σq]=[σ~1,…,-,…,σ~q]=0otherwisek_n(\sigma, \tilde{\sigma}) = \begin{cases} \neq 0 & \exists u, \text{ s.t. } [\sigma_1, \dots, \text{-}, \dots, \sigma_q] = [\tilde{\sigma}_1, \dots, \text{-}, \dots, \tilde{\sigma}_q] \\\\ = 0 & \text{otherwise} \end{cases}

  3. Consider its reversal: μn(σ)=∑σ~knrev(σ,σ~)μn+1(σ~)whereknrev(σ,σ~)=kn(σ~,σ)∑σ^kn(σ~,σ^)μn(σ^)μn(σ)\mu_{n}(\sigma) = \sum_{\tilde{\sigma}} k_n^{\text{rev}}(\sigma, \tilde{\sigma}) \mu_{n+1}(\tilde{\sigma}) \quad \text{where} \quad k_n^{\text{rev}}(\sigma, \tilde{\sigma}) = \frac{k_n(\tilde{\sigma}, \sigma)}{\sum_{\hat{\sigma}} k_n(\tilde{\sigma}, \hat{\sigma}) \frac{\mu_n(\hat{\sigma})}{\mu_n(\sigma)}}

  4. Make model estimating for μn(σ~)μn(σ)\frac{\mu_n(\tilde{\sigma})}{\mu_n(\sigma)} under the assumption: μn(σ~)\mu_n(\tilde{\sigma}) and μn(σ)\mu_n(\sigma) differ at only one coordinate nn, which gives μn(σ~)μn(σ)=μn(σ~n∣σ1,…,-,…,σq)μn(σn∣σ1,…,-,…,σq)\frac{\mu_n(\tilde{\sigma})}{\mu_n(\sigma)} = \frac{\mu_n(\tilde{\sigma}_n \mid \sigma_1, \dots, \text{-}, \dots, \sigma_q)}{\mu_n(\sigma_n \mid \sigma_1, \dots, \text{-}, \dots, \sigma_q)} and the model approximating μn(σ~n∣σ1,…,-,…,σq)∝exp⁡(Hu(σ))≈exp⁡(H~u(σ;w))\mu_n(\tilde{\sigma}_n \mid \sigma_1, \dots, \text{-}, \dots, \sigma_q) \propto \exp(H_u(\sigma)) \approx \exp(\tilde{H}_u(\sigma; w))

  5. Run the inference based on reverse process with H~u(σ;w)\tilde{H}_u(\sigma; w) to mimic sampling from the target distribution μ0\mu_0.

The training model is designed as: Hu~(σ;w)=∑s=1qΦs(σu)NNθ(σn∣σ1,…,-,…,σq)\tilde{H_u}(\sigma; w) = \sum_{s=1}^q \Phi_s(\sigma_u) NN_\theta(\sigma_n \mid \sigma_1, \dots, \text{-}, \dots, \sigma_q) where Φs(σu)=1−1q\Phi_s(\sigma_u) = 1 - \frac{1}{q} if σu=s\sigma_u = s, and −1q-\frac{1}{q} otherwise.

They validate their proposals in the numerical experiments.

Contributions

The contributions of this paper are

  • formulation of the discrete diffusion model based on modeling energy function of single-site conditional distributions given the remaining coordinates,
  • reformulation of the autoregressive models from this viewpoint.

Validation

The authors validate the formulation by running some numerical experiments:

  • Test case1: Binary data:
    • Ising model
      • the NeuRISE-based denoising estimator shows the sharpest decay in total variation (TV) distance with increase in sample size, and performs better than the GGM based estimator of Varma et al. (2024), the SEDD approach proposed in Lou et al. (2024) and the discrete flow matching method of Gat et al. (2024).
    • MNIST
      • the NeuRISE based learning of conditionals achieve the lowest MMD and cross-correlation error.
    • D-wave dataset
      • Results in Table 1b show the advantage of NeurISE Diffusion in all metrics.
  • Test case2: Multi-valued data:
    • qq-state potts model
      • the TV error decreases as the number of training samples are increased.
  • Test case3: Multi-valued data:
    • quantum tomography of GHZ state
      • the cross-correlation error goes down significantly after 10410^4 samples, indicating that the model is able to learn a faithful generative model for this quantum state.
Critical strengths
  • While I do not believe this goes beyond the framework of standard discrete diffusion models, there may be some value in reframing it as the learning of a local Hamiltonian.
Critical weaknesses
  • The claim that "numerical experiments demonstrate the superiority of NeuroISE diffusion models across all metrics" seems too good to be true. We can't identify any decisive reason why such remarkably favorable results would emerge.
  • As previously noted, the theoretical component of this paper, i.e., applying round-robin noise modeling and its inverse, appears to be a standard approach in discrete diffusion models. Replacing this with energy function learning merely involves taking the logarithm of the standard model, suggesting little novelty in this aspect.
Questions for authors

Question regarding the aforementioned weakness

  • Does simply modifying the conventional discrete diffusion model configuration to employ learned local energy function produce such dramatic performance improvements as the author emphasized? Additionally, while hyperparameters were reportedly automatically tuned, was there proper trial-and-error conducted regarding the neural network architecture and other model design aspects?
  • If this formulation is novel, what specifically makes it a new idea compared to existing methods?

General questions

  • Φs(r)\Phi_s(r) is defined for pairs s,r∈Σs, r \in \Sigma, but how are the vector-valued components like Φ1(r)\Phi_1(\mathbf{r}) defined?

  • The rationale for introducing Φs\Phi_s remains unclear for me. Ultimately, Φs\Phi_s only produces differences of nearly constant multiplicative factors within the portion defining Hu~(σ;w)\tilde{H_u}(\sigma; w), making that portion unnecessary and better handled by NNθ(σ−u)sNN_\theta(\sigma_{-u})_s. Wouldn't that make the original portion redundant?

  • At various points, σu\sigma_u and σn\sigma_n appeared in similar contexts, but is there a deliberate distinction between them? If so, what is its meaning?

  • The authors write as follows:

    In order to implement the reverse dynamics, we need to estimate the ratio μn(σ~)μn(σ)\frac{\mu_n(\tilde{\sigma})}{\mu_n(\sigma)} from samples of the forward process. Suppose σ,σ~∈Σq\sigma, \tilde{\sigma} \in \Sigma^q differ at only one coordinate nn, i.e., σ~i=σi\tilde{\sigma}_i = \sigma_i for all i≠ni \neq n, σ~n≠σn\tilde{\sigma}_n \neq \sigma_n. Then, for any distribution μn\mu_n over Σq\Sigma^q, we have:

    The part seems to derive from the forward process in a round-robin fashion, but does it need to be assumed?

  • At line 281, the final character before the closing bracket should be a comma instead of a period?

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

This paper studies a discrete denoising diffusion framework, which is a combination of two existing methods, Neural Interaction Screening Estimator (NeurISE) and round-robin noising diffusion. The framework is evaluated on MLPs and several datasets including synthetic datasets and MNIST.

Critical strengths
  • The connection between conditional estimation and generation is clearly stated.
Critical weaknesses
  • The novelty of this work seems limited, since both round-robin noising and NeurISE are inherited from previous works. The main contribution of this work is the integration of the two parts, but the paper does not clearly identify new technical insights or challenges resolved by this combination.
  • The empirical evaluations are restricted to MLPs. It remains unclear whether the reported advantages of the proposed algorithm persist with other network architectures or other datasets.
Questions for authors
  • Why are some of the reported total variation distances in Figures 3.(a) and 4.(a) greater than 1?
Reviewer_X1cLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a discrete diffusion model that parameterizes the reverse process through single-site conditionals. The authors employ round-robin forward noising, where a noise parameter controls the randomization of the forward process. Later, the authors propose use the NeurISE loss to train a neural network, which is also used for sampling in the reverse process. Finally, the authors evaluate their method on several scientific datasets and MNIST.

Critical strengths

The connection with physics is interesting and opens the door to a physical interpretation of discrete diffusion models.

Critical weaknesses
  • The experimental evidence is thin: the authors ignore stronger baselines like [1,2]. Additionally, the datasets are small-scale and do not constitute evidence for the scalability of the method. The presentation of the results is suboptimal: in figure 1 the logarithmic scale distorts the unfavourable gap at low training sizes and inflates the smaller difference at higher sample sizes. A table with raw numbers would be easier to read. Several experimental details are also missing, like the transition type and the number of function evaluations the authors used for SEDD and D3PM.
  • The novelty of the proposed method is weak: as the authors acknowledge the round-robin forward process was already used in GGM [3] and the training loss is the one from NeurISE [4]. In the current version, however, the presentation makes it ambiguous which parts are original. The authors should create a preliminaries section to separate background and original contributions. Likewise, the reverse kernel and the ratio parameterization are presented as contributions while they are standard results in the discrete diffusion model literature [5,6]. The paper could be improved by simplifying the derivations of known results while expanding on the theoretical properties and empirical evidence.
  • The single-site update of the proposed method does not allow parallel decoding, which is a decisive advantage in choosing discrete diffusion models over autoregressive models in practical applications.

[1] Sahoo et al. ; The diffusion duality

[2] Shi et al. ; Simplified and Generalized Masked Diffusion for Discrete Data

[3] Varma et al. ; Glauber generative model: Discrete diffusion models via binary classification.

[4] Jayakumar et al. ; Learning of discrete graphical models with neural networks

[5] Lou et al. ; Discrete diffusion modeling by estimating the ratios of the data distribution

[6] Meng et al. ; Concrete score matching: Generalized score matching for discrete data

Questions for authors
  • Can the authors provide empirical evidence that the proposed method scales to more complex datasets, such as CIFAR-10? I understand that it might be tricky to run new experiments at rebuttal time but the lack of empirical evidence is a big drawback in the current paper.
  • Why does the training time of NeurISE remain almost constant in figure 1c? Can the authors provide the number of steps they used for training, whether they used early stopping for all methods, and all the details necessary to understand figure 1c?
  • Why does the proposed method perform worse than D3PM in figure 1a in the low-sample regime?

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

Related papers

Discussion (0)

Sign in to comment.