Pith. sign in

REVIEW 3 major objections 7 minor 37 references

Conditional Distribution Quantization in Machine Learning

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Minimizing the conditional quantization loss is equivalent to minimizing the expected squared 2-Wasserstein distance between $\mathcal{L}(Y\mid X)$ and an $n$-point quantized approximation, and an optimal quantizer attains the best…

desk verdict A clean conditional quantization identity with a real gap between the Wasserstein guarantee and the classifier-weighted mixture the model actually outputs. read the letter →

arxiv 2502.07151 v2 pith:BSAV267N submitted 2025-02-11 cs.LG

Blaise Delattre , Sylvain Delattre , Alexandre Vérine , Alexandre Allauzen This is my paper · ORCID
classification cs.LG
keywords conditionalquantizationcompetitivelearningvectorWassersteindistancewinner-takes-alluncertaintyquantificationmultimodalregressionimageinpaintinggenerativemodels
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

This paper claims that a conditional distribution $\mathcal{L}(Y \mid X)$ can be approximated by $n$ learnable functions $f_1(X),\dots,f_n(X)$ whose outputs define a quantized, finite-support conditional law, and that the distortion $\Delta_n(f)=\mathbb{E}[\min_i |Y-f_i(X)|^2]$ is exactly the expected squared 2-Wasserstein distance between the true conditional law and that quantized approximation. That identification, Theorem 4.1, turns a competitive-learning objective into a principled way to represent multimodality: instead of a single conditional mean, the model outputs several representative points together with estimated probabilities. Theorem 4.2 adds that an optimal choice of the $n$ functions exists and attains, almost surely, the best approximation of the conditional law among all distributions supported on $n$ points. The paper's algorithm, Conditional Competitive Learning Vector Quantization (CCLVQ), trains expert networks with a winner-takes-all update and a small classifier for the cell probabilities, and demonstrates the idea on synthetic regression, MNIST and CelebA inpainting and denoising, normalizing flows, and GAN diversity.

What carries the argument

The central object is the conditional $n$-point quantizer $f=(f_1,\dots,f_n):E\to(\mathbb{R}^d)^n$ with distortion $\Delta_n(f)=\mathbb{E}[\min_{1\le i\le n}|Y-f_i(X)|^2]$. For each $x$, the closest-neighbor projection $\pi_{f(x)}$ induces the quantized conditional law $\widehat{Q}_f(x,\cdot)=\sum_{i=1}^n Q(x,\{y:\pi_{f(x)}(y)=f_i(x)\})\,\delta_{f_i(x)}$. The load-bearing identity of Theorem 4.1 is the conditional version of the classical quantization identity: with probability one, $\mathbb{E}[\min_i|Y-f_i(X)|^2\mid X]=W_2^2(Q(X,\cdot),\widehat{Q}_f(X,\cdot))$, so $\Delta_n(f)=\mathbb{E}[W_2^2(Q(X,\cdot),\widehat{Q}_f(X,\cdot))]$. The mechanism that minimizes this loss is CCLVQ: assign each sample to the expert minimizing the squared distance, update only that expert by gradient descent, and train a classifier $h$ on the assignments to estimate the cell probabilities $Q(x,\{y:\pi_{f(x)}(y)=f_i(x)\})$).

What would settle it

Train two expert networks on a symmetric two-mode conditional law, say $Y=X+100$ with probability $1/2$ and $Y=X-100$ with probability $1/2$, then compare the converged CCLVQ loss to the expected squared Wasserstein distance computed from the empirical conditional law; if the selected-expert update is biased, a residual gap that persists across learning rates and seeds would show the optimization claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is Theorem 4.1: for every conditional quantizer $f=(f_1,\ldots,f_n):E\to(\mathbb{R}^d)^n$, the quantization loss $\Delta_n(f)=\mathbb{E}[\min_{1\le i\le n}|Y-f_i(X)|^2]$ satisfies $\Delta_n(f)=\mathbb{E}[W_2^2(Q(X,\cdot),\widehat{Q}_f(X,\cdot))]$, where $Q(X,\cdot)=\mathcal{L}(Y\mid X)$ and $\widehat{Q}_f(X,\cdot)$ is the law of the closest-neighbor projection of $Y$ onto $\{f_1(X),\dots,f_n(X)\}$ given $X$. Thus minimizing $\Delta_n(f)$ is equivalent to reducing the expected squared Wasserstein distance to the true conditional law. Theorem 4.2 then proves that an optimal quantizer $f^*$ exists and that any optimal quantizer attains, with probability one, $W_2(Q(X,\cdot),\widehat{Q}_{f^*}(X,\cdot))=\min_{\nu\in\mathcal{P}_n(\mathbb{R}^d)}W_2(Q(X,\cdot),\nu)$, the best approximation of the conditional law by any distribution supported on at most $n$ points.

Load-bearing premise

The Wasserstein guarantee depends on the assumption that updating only the closest expert on each sample is an unbiased gradient step on the distortion when the experts are neural networks; this is proved for ordinary finite-dimensional quantizers but not for the neural case, and if it fails the trained model may not actually approach a conditional quantizer.

Editorial extensions

If this is right

  • A model trained to near-optimal $\Delta_n(f)$ outputs $n$ representative points for each input together with estimated probabilities, so the whole conditional law $\mathcal{L}(Y\mid X)$, including its multimodality, is available for downstream decisions instead of a single conditional mean.
  • Because the approximation is measured in Wasserstein distance, the uncertainty reported by CCLVQ is structured rather than per-pixel: it accounts for spatial correlations and offers several plausible reconstructions, which is what inpainting and denoising require.
  • Adding experts by splitting an existing one decreases $\Delta_n$, and the reported normalized entropies near $1$ on CIFAR-10 and CelebA indicate the added experts are used roughly uniformly rather than collapsing.
  • The same objective transfers to generative models: normalizing flows with $n$ expert branches produce fewer out-of-distribution samples, and a mixture-of-experts BigGAN improves FID and recall while keeping precision fixed.

Reading between the lines

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

  • Because Theorem 4.1 is written for squared Euclidean distance, the generalization of CCLVQ to arbitrary losses $\ell$ is not covered by the Wasserstein guarantee; only losses proportional to squared distance inherit that interpretation, so other losses need their own analysis.
  • A testable extension is to compare CCLVQ with mixture density networks on the same conditional laws: the Wasserstein identity predicts the experts' Voronoi cells at convergence should match the modes learned by an MDN with the same $n$, and the two output measures should approach each other as $n$ grows.
  • The theory concerns arbitrary functions $f_i$, so the gap between a trained neural quantizer and the ideal one is controlled by network capacity; measuring $\mathbb{E}[W_2^2]$ on held-out conditional laws while scaling capacity and $n$ should show a monotone decrease toward the $n$-point quantization error.
  • The classifier is trained on hard expert assignments, which can bias the estimated cell weights near boundaries; a soft or EM-like assignment would estimate the weights in $\widehat{Q}_f$ with less bias, and the theory suggests those weights are the correct uncertainty readout.
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

3 major / 7 minor

Summary. The paper proposes Conditional Competitive Learning Vector Quantization (CCLVQ), which approximates a conditional distribution L(Y | X) by n input-dependent experts f_i(X). The main theoretical results are Theorem 4.1, showing that the distortion Delta_n(f) = E[min_i |Y - f_i(X)|^2] equals the expected squared Wasserstein distance between the true conditional law Q(X, ·) and the Voronoi-weighted quantized measure \hat{Q}_f(X, ·), and Theorem 4.2, asserting the existence of an optimal conditional quantizer and its pointwise optimality. An algorithm is then presented that alternates winner-takes-all assignment of samples to experts with gradient updates of the selected expert, and trains a classifier h to estimate the expert weights. Applications to multimodal regression, image inpainting/denoising, normalizing flows, and GANs are demonstrated.

Significance. If the gaps identified below are addressed, this paper offers a principled way to summarize multimodal conditional laws with a finite set of input-dependent points, with a clean equivalence between the training loss and the expected squared Wasserstein distance. The core identities in Theorem 4.1 are correct under the stated integrability assumption and are a useful conditional extension of classical quantization theory. The experimental results, while largely qualitative for imaging tasks, support the potential of the method for uncertainty quantification and diversity enhancement. The paper does not currently ship reproducible code, and the quantitative evaluation is limited (two seeds, no statistical significance tests).

major comments (3)
  1. [Theorem 4.1 and Section 4.1 (Algorithm 1)] The theoretical guarantee of Theorem 4.1 concerns the Voronoi-weighted measure \hat{Q}_f(X, ·), whose atoms carry the true conditional cell probabilities Q(x, {y : π_f(x)(y) = f_i(x)}). The model actually deployed in Algorithm 1 outputs the classifier-weighted mixture \sum_i h_i(x) δ_{f_i(x)}, and no analogue of Theorem 4.1 is proved for these learned weights. By Proposition 3.1, the optimal weights on a fixed support {f_i(x)} are exactly the Voronoi cell probabilities, so any deviation of h from them can strictly increase W2 to Q(x, ·). A near-minimizer of the CCLVQ loss therefore need not yield a low-W2 approximation of the distribution the method returns; this gap is visible in Figure 3, where the estimated weights deviate from the true mode probabilities in low-density regions. Please either prove that the cross-entropy-trained h consistently estimates the Voronoi cell probabilities and state a bound on the resulting W2 error, or explicitly restrict the theoretical claims to the idealized Voronoi-weighted measure.
  2. [Section 4.1 (CCLVQ update)] The paper claims the CCLVQ update is 'closely linked to gradient descent,' but Proposition 3.3 is proved only for the finite-dimensional quantizer α, not for parameterized experts f_i(·; θ_i). No conditional analogue establishing that the winner-takes-all update is an unbiased stochastic gradient of Δ_n(f_θ) is provided. For neural parameterizations, the update takes the assignment I_f(X)(Y) as fixed and differentiates only the selected expert's output, omitting the dependence of the assignment on θ; it is therefore not generally an unbiased gradient of Δ_n(f_θ). Since Theorems 4.1–4.2 describe global minimizers of Δ_n, the Wasserstein guarantee may not transfer to the actual trained model. Please state explicitly the sense in which the algorithm optimizes Δ_n (e.g., as a hard-EM or majorization-minimization heuristic) and what conditions would be needed for its fixed points to approach minimizers.
  3. [Appendix A (proof of Theorem 4.2)] The proof constructs f*(x) pointwise as an optimal quantizer for Q(x, ·), but it does not establish that the resulting map x ↦ f*(x) is measurable. Measurability is required for f* to be an admissible element of the domain of Δ_n and for the expectations in the proof to be defined. Please add a measurable selection argument (e.g., apply a measurable selection theorem to the optimal-quantizer correspondence) or otherwise justify measurability under the stated assumptions.
minor comments (7)
  1. [Appendix A, first displayed equation] In the proof of Theorem 4.2, the line 'for any f, one has Δ_n(f*) = E[min_i |Y - f_i(X)|^2]' should read f*_i(X) instead of f_i(X) in the minimum; the notation currently suggests the inequality is being proved with the wrong quantizer.
  2. [Throughout] The text contains several typos: 'On has' in Theorem 4.1(2), 'F or every' in Proposition 3.1, 'Y oseph Linde' in the references, and the running header reads 'Conditional Distribution Quantification' on pages 2–3.
  3. [Figure 4 and Section 5.2] The caption of Figure 4 says 'The first column X represents the base image' while the figure legend shows the first column as ground truth Y and the second as masked input X; please align the caption with the figure.
  4. [Section 5.3] The normalizing-flow experiment uses a likelihood-based assignment rather than the squared Euclidean loss; the theoretical guarantee in Theorem 4.1 is specific to the quadratic loss and the Wasserstein-2 metric, so the connection to the Wasserstein statement is not justified for this experiment without an additional result for general loss functions.
  5. [Section 4.1] The update rule displayed after 'One iteration associated with one data point (X,Y)' has a mismatched parenthesis: the indicator should multiply the gradient inside the vector, i.e., (1_{I_f(X)(Y)=i} ∇_{θ_i} ||f_i(X,θ_i) - Y||^2)_{1≤i≤n} with a closing bracket for the norm.
  6. [Figure 1] The 'Choice of n' experiment reports only the training loss; reporting the test distortion would better support the claim that adding experts improves generalization.
  7. [Algorithm 1, step 9] Tie-breaking in the argmin assignment is not specified; as in Proposition 3.3, a zero-tie-probability condition is needed for the update to be well-defined and for gradient statements to hold.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central Wasserstein identity is a proved theorem, not a fitted quantity or a self-citation.

full rationale

The paper's main theoretical claim, Theorem 4.1, states Delta_n(f) = E[W2(Q(X,·), Qhat_f(X,·))^2]. This is not circular: Qhat_f is defined through the Voronoi cells of f and the true conditional law Q, and the equality is proved by applying Proposition 3.1 conditionally on X. Proposition 3.1 itself is a standard quantization result with an independent proof given in the paper; the authors cite Graf-Luschgy and Pagès only for background existence results. No fitted parameter is renamed as a prediction: the expert functions f and classifier weights h are estimated from data, but the theorems do not depend on the numerical values of these estimates. The self-citations (Verine et al., 2023a,b) are used only to motivate why normalizing flows and GANs might benefit from a multi-expert approach; they are not load-bearing premises of Theorems 4.1 or 4.2. The review's concern about the classifier-weighted mixture versus the Voronoi-weighted measure bQ_f is a genuine approximation-gap issue in the algorithm-to-theory transfer, but it is not circularity: the learned mixture is an estimator of bQ_f, and no equation in the paper reduces the advertised guarantee to the fitted weights. The paper is self-contained at the level of its mathematical claims, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No invented physical or latent entities are introduced. The framework depends on standard quantization theory, finite-second-moment assumptions, an unproved measurable selection step, and an assumption that the WTA update behaves as gradient descent. The number of experts and splitting schedule are free hyperparameters chosen empirically.

free parameters (3)
  • number of experts n = n = 3 in Figure 1, 10 in Appendix B.1/B.3, 1/2/4 for BigGAN
    Chosen empirically per task; the paper states the choice of n is empirical and depends on the task.
  • expert splitting schedule = every 400 epochs in Figure 1, one final epoch for CelebA-HQ in Appendix B.3
    The timing of when experts are duplicated is a heuristic design choice that affects training dynamics.
  • assignment noise scale = not reported
    Appendix B.3 adds small Gaussian noise to the loss to avoid winner-takes-all collapse, but the noise magnitude is not specified.
assumptions (5)
  • standard math For every probability distribution with finite second moment, an optimal n-point quantizer exists.
    Used in the proof of Theorem 4.2 and cited to Graf and Luschgy 2000 and Pages 2015.
  • domain assumption The conditional law Q(X,·) has finite second moment almost surely and some f with Delta_n(f) < infinity exists.
    Stated in Section 4 as the integrability assumption; it makes the Wasserstein distances finite.
  • ad hoc to paper A measurable selection of the pointwise optimal quantizer f*(x) exists.
    Appendix A constructs f* pointwise for each x but does not prove the resulting map is measurable; without a measurable selection theorem f*(X) may not be a valid random variable.
  • ad hoc to paper The CCLVQ winner-takes-all update is an unbiased stochastic gradient step for Delta_n.
    Section 4.1 says the algorithm is 'closely linked to gradient descent', but the conditional analogue of Proposition 3.3 and the tie condition are not proved for neural parameterizations.
  • domain assumption Nearest-neighbor assignment ties have probability zero at training iterates.
    Required for differentiability of the distortion as in Proposition 3.3, condition (3); not verified for learned experts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conditional Distribution Quantization in Machine Learning." pith.science (2026). https://pith.science/paper/BSAV267N

@misc{pith2026250207151,
  author       = {Pith},
  title        = {Pith review of: Conditional Distribution Quantization in Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BSAV267N}},
  note         = {Machine review of arXiv:2502.07151}
}
read the original abstract

Conditional expectation \mathbb{E}(Y \mid X) often fails to capture the complexity of multimodal conditional distributions \mathcal{L}(Y \mid X). To address this, we propose using n-point conditional quantizations--functional mappings of X that are learnable via gradient descent--to approximate \mathcal{L}(Y \mid X). This approach adapts Competitive Learning Vector Quantization (CLVQ), tailored for conditional distributions. It goes beyond single-valued predictions by providing multiple representative points that better reflect multimodal structures. It enables the approximation of the true conditional law in the Wasserstein distance. The resulting framework is theoretically grounded and useful for uncertainty quantification and multimodal data generation tasks. For example, in computer vision inpainting tasks, multiple plausible reconstructions may exist for the same partially observed input image X. We demonstrate the effectiveness of our approach through experiments on synthetic and real-world datasets.

Figures

Figures reproduced from arXiv: 2502.07151 by the authors.

Figure 1
Figure 1. CCLVQ training loss on the MNIST dataset for an in￾painting task, showing the effect of adding new experts using the splitting strategy every 400 epochs. where, for α ∈ (R d ) n and y in R d , Iα(y) ∈ {1, . . . , n} satisfies ℓ(y, αIα(y)) = min 1≤i≤n ℓ(y, αi). The algorithm selectively updates only the parameters of the function that minimizes the distortion for a given sam￾ple, iteratively refining the functions to… view at source ↗
Figure 2
Figure 2. Visualization of synthetic data showing true data points Y (in grey) and predicted values from every expert (in colors). Each expert captures a specific mode of conditional distribution. 5. Examples of applications 5.1. Multi-value Regression on synthetic data We aim to demonstrate CCLVQ’s applications through multi￾value regression, multi-modal reconstruction, and genera￾tion experiments. Upon acceptance, the code … view at source ↗
Figure 5
Figure 5. Multi-modal denoising results on MNIST. Each row shows the ground truth Y , noised input X, and reconstructions from different experts, followed by the classifier’s weights as a measure of uncertainty [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Multi-modal inpainting results on MNIST. Each row shows the ground truth Y , masked input X, and reconstructions from different experts, followed by the classifier’s weights as a measure of uncertainty. 5.2. Multi-modal Reconstruction with Uncertainty Quantification MN…
Figure 7
Figure 7. Figure 7: Samples generated by RealNVP trained on a mixture of 2D Gaussians. As the number of experts increases, the model generates fewer out-of-distribution samples. Each expert is represented by a unique color. Metrics CIFAR-10 CelebA n 1 2 4 2∗ 1 2 FID (↓) 8.21 ± 0.72 7.34 ±…
Figure 8
Figure 8. Figure 8: Inpainting results on MNIST for a unique model case. Each row shows the ground truth Y , masked input X, and reconstructions from the model [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Denoising results on MNIST for a unique model case. Each row shows the ground truth Y , noised input X, and reconstructions from the model. B. Experimental details B.1. Inpainter Architecture and Training Details on MNIST On the Experiment B.2, the Inpainter model empl…
Figure 10
Figure 10. Figure 10: More inpainting results on the CelebA-HQ dataset using CCLVQ. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Generated CIFAR-10 samples from BigGAN models with n = 1 and n = 2 experts. (a) n = 1 (b) n = 2, Model 1 (c) n = 2, Model 2 (d) n = 2, Mixture [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Generated CelebA samples from BigGAN models with n = 1 and n = 2 experts. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Image-to-image regression with distribution-free uncertainty quantification and applications in imaging

    Anastasios N Angelopoulos, Amit P Kohli, Stephen Bates, Michael I Jordan, Jitendra Malik, Thayer Alshaabi, Srigokul Upadhyayula, and Yaniv Romano. Image-to-image regression with distribution-free uncertainty quantification and applications in imaging. In International Conference on Machine Learning, 2022

    work page 2022

  3. [3]

    pca GAN : Improving posterior-sampling c GAN s via principal component regularization

    Matthew C Bendel, Rizwan Ahmad, and Philip Schniter. pca GAN : Improving posterior-sampling c GAN s via principal component regularization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  4. [4]

    Christopher M. Bishop. Mixture density networks. Technical report, Neural Computing Research Group, Aston University, 1994

    work page 1994

  5. [5]

    Weight uncertainty in neural networks

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. In International Conference on Machine Learning, pages 1613--1622. PMLR, 2015

    work page 2015

  6. [6]

    About the multidimensional competitive learning vector quantization algorithm with constant gain

    Catherine Bouton and Gilles Pag\`es. About the multidimensional competitive learning vector quantization algorithm with constant gain. Ann. Appl. Probab., 7 0 (3): 0 679--710, 1997. ISSN 1050-5164,2168-8737. doi:10.1214/aoap/1034801249. URL https://doi.org/10.1214/aoap/1034801249

    work page arXiv 1997

  7. [7]

    Large scale GAN training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=B1xsqj09Fm

    2019

  8. [8]

    Relaxing bijectivity constraints with continuously indexed normalising flows

    Rob Cornish, Anthony Caterini, George Deligiannidis, and Arnaud Doucet. Relaxing bijectivity constraints with continuously indexed normalising flows. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2133--2143. PMLR, 13--18 Jul 202...

    work page 2020

Show all 37 references
  1. [9]

    Density estimation using real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP . In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=HkpbnH9lx

    2017

  2. [10]

    U-net: deep learning for cell counting, detection, and morphometry

    Thorsten Falk, Dominic Mai, Robert Bensch, Özgün Çiçek, Ahmed Abdulkadir, Yassine Marrakchi, Anton Böhm, Jan Deubner, Zoe Jäckel, Katharina Seiwald, et al. U-net: deep learning for cell counting, detection, and morphometry. Nature Methods, 2019

    2019

  3. [11]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, pages 1050--1059. PMLR, 2016

    2016

  4. [12]

    Foundations of quantization for probability distributions, volume 1730 of Lecture Notes in Mathematics

    Siegfried Graf and Harald Luschgy. Foundations of quantization for probability distributions, volume 1730 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2000. ISBN 3-540-67394-6. doi:10.1007/BFb0103945. URL https://doi.org/10.1007/BFb0103945

    doi

  5. [13]

    Multiple choice learning: Learning to produce multiple structured outputs

    Alejandro Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. In Advances in Neural Information Processing Systems (NeurIPS), 2012

    2012

  6. [14]

    GANs Trained by a Two Time - Scale Update Rule Converge to a Local Nash Equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs Trained by a Two Time - Scale Update Rule Converge to a Local Nash Equilibrium . In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017. URL...

    2017

  7. [15]

    What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems (NeurIPS), 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  8. [16]

    Learning vector quantization for pattern recognition

    Teuvo Kohonen. Learning vector quantization for pattern recognition. In Proceedings of the International Conference on Neural Networks, 1995

    1995

  9. [17]

    Conformal prediction masks: Visualizing uncertainty in medical imaging

    Gilad Kutiel, Regev Cohen, Michael Elad, Daniel Freedman, and Ehud Rivlin. Conformal prediction masks: Visualizing uncertainty in medical imaging. In Proceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  10. [18]

    Improved Precision and Recall Metric for Assessing Generative Models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models . In 33rd Conference on Neural Information Processing Systems ( NeurIPS 2019), Vancouver , Canada . , October 2019. arXiv: 1904.06991

    2019 arXiv

  11. [19]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, pages 6402--6413, 2017

    2017

  12. [20]

    Confident multiple choice learning

    Kimin Lee, Changho Hwang, KyoungSoo Park, and Jinwoo Shin. Confident multiple choice learning. In Proceedings of the 34th International Conference on Machine Learning, 2017

    2017

  13. [21]

    Stochastic multiple choice learning for training diverse deep ensembles

    Stefan Lee, Senthil Purushwalkam, Michael Cogswell, Varun Ranjan, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems (NeurIPS), 2016

    2016

  14. [22]

    Resilient multiple choice learning: A learned scoring scheme with application to audio scene analysis

    Victor Letzelter, Mathieu Fontaine, Mickael Chen, Patrick Perez, Slim Essid, and Ga \"e l Richard. Resilient multiple choice learning: A learned scoring scheme with application to audio scene analysis. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  15. [23]

    Winner-takes-all learners are geometry-aware conditional density estimators

    Victor Letzelter, David Perera, C \'e dric Rommel, Mathieu Fontaine, Slim Essid, Ga \"e l Richard, and Patrick P \'e rez. Winner-takes-all learners are geometry-aware conditional density estimators. In Proceedings of the 41st International Conference on Machine Learning, 2024

    2024

  16. [24]

    Implicit maximum likelihood estimation, 2019

    Ke Li and Jitendra Malik. Implicit maximum likelihood estimation, 2019. URL https://openreview.net/forum?id=rygunsAqYQ

    2019

  17. [25]

    An algorithm for vector quantizer design

    Yoseph Linde, Andres Buzo, and Robert M Gray. An algorithm for vector quantizer design. IEEE Transactions on Communications, 1980

    1980

  18. [26]

    Least squares quantization in pcm

    Stuart P Lloyd. Least squares quantization in pcm. IEEE Transactions on Information Theory, 1982

    1982

  19. [27]

    On the posterior distribution in denoising: Application to uncertainty quantification

    Hila Manor and Tomer Michaeli. On the posterior distribution in denoising: Application to uncertainty quantification. In Proceedings of the International Conference on Learning Representations (ICLR). ICLR, 2024

    2024

  20. [28]

    Uncertainty quantification via neural posterior principal components

    Elias Nehme, Omer Yair, and Tomer Michaeli. Uncertainty quantification via neural posterior principal components. In Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS). NeurIPS, 2023

    2023

  21. [29]

    Introduction to vector quantization and its applications for numerics

    Gilles Pagès. Introduction to vector quantization and its applications for numerics. ESAIM: Proceedings and Surveys, 48 0 (1): 0 29--79, 2015. doi:https://doi.org/10.1051/proc/201448002. URL http://www.sciengine.com/publisher/EDP Sciences/journal/ESAIM: Proceedings and Surveys...

    2015

  22. [30]

    a henb \

    Deepak Pathak, Philipp Kr \"a henb \"u hl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016

    2016

  23. [31]

    CHIMLE : Conditional hierarchical IMLE

    Shichong Peng, Seyed Alireza Moazenipourasil, and Ke Li. CHIMLE : Conditional hierarchical IMLE . In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022

    2022

  24. [32]

    Annealed multiple choice learning: Overcoming limitations of winner-takes-all with annealing

    David Perera, Victor Letzelter, Theo Mariotte, Adrien Cortes, Mickael Chen, Slim Essid, and Ga \"e l Richard. Annealed multiple choice learning: Overcoming limitations of winner-takes-all with annealing. In The Thirty-eighth Annual Conference on Neural Information Processing S...

    2024

  25. [33]

    Computational Optimal Transport: With Applications to Data Science

    Gabriel Peyr \'e and Marco Cuturi. Computational Optimal Transport: With Applications to Data Science. Foundations and Trends in Machine Learning, 2019

    2019

  26. [34]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention--MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part...

    2015

  27. [35]

    Christian Rupprecht, Iro Laina, Robert DiPietro, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D. Hager. Learning in an uncertain world: Representing ambiguity through multiple hypotheses. In Proceedings of the IEEE International Conference on Computer Vision (...

    2017

  28. [36]

    Precision-recall divergence optimization for generative modeling with GAN s and normalizing flows

    Alexandre Verine, benjamin negrevergne, Muni Sreenivas Pydi, and Yann Chevaleyre. Precision-recall divergence optimization for generative modeling with GAN s and normalizing flows. In Thirty-seventh Conference on Neural Information Processing Systems, 2023 a . URL https://open...

    2023

  29. [37]

    On the expressivity of bi-lipschitz normalizing flows

    Alexandre Verine, Benjamin Negrevergne, Yann Chevaleyre, and Fabrice Rossi. On the expressivity of bi-lipschitz normalizing flows. In Emtiyaz Khan and Mehmet Gonen, editors, Proceedings of The 14th Asian Conference on Machine Learning, volume 189 of Proceedings of Machine Lear...

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.