acceptodds
Under review as a conference paper at ICLR 2027

Availability and Predictivity: Improving Coreset Training at Few Images Per Class with Covariance Jitter

Abstract

Coreset selection aims to replace a large training set with a small subset while preserving downstream performance. We show that models trained on just a few images per class (IPC) rely on features along the data’s high-variance directions, even when features along the low-variance directions are more predictive. This motivates our method, Coreset Covariance Jitter (CCJ), a simple, low-cost training augmentation applicable to any existing coreset selection method. CCJ adds noise to each selected image at every training step, scaled along each direction by how much images of the same class vary along it in the full training set. We further increase performance by adding a Quadratic Discriminant Analysis (QDA) at test time. On CIFAR-10, CIFAR-100 and ImageNet-1K, CCJ+QDA improves all evaluated coreset selection methods when running at 10 images per class.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
coreset selection, dataset pruning, feature learning, simplicity bias

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dXeMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates why models trained on small coresets tend to favor high-variance features over potentially more predictive low-variance features. It introduces Coreset Covariance Jitter (CCJ), which perturbs training images according to estimated within-class feature variances, and combines it with a QDA classifier at inference to recover complementary class information. The paper provides a theoretical analysis of feature availability and predictivity, supported by experiments on a controlled synthetic task. Extensive evaluations on CIFAR-100 and ImageNet-1K show consistent improvements across coreset selection methods at 10 images per class, with additional validation across model architectures, label-free settings, and dataset distillation. The method adds little computational or storage overhead, though its benefits are less consistent at higher coreset budgets.

Critical strengths
  1. Strong empirical performance with minimal overhead. CCJ consistently improves a wide range of coreset selection methods at IPC-10 on CIFAR-100 and ImageNet-1K. Its low computational and storage costs make it a practical enhancement to existing pipelines rather than a replacement for them.

  2. A useful theoretical perspective with a simple, generalizable method. The paper connects feature availability and predictivity to explain why models trained on small coresets favor high-variance features. CCJ translates this insight into a lightweight augmentation, while QDA provides complementary information at inference time. The combination is validated across multiple architectures and label-free settings, strengthening the practical relevance of the contribution.

Critical weaknesses

Dependence on full-dataset statistics limits the practical interpretation of coreset compression. Both CCJ and QDA require statistics computed from the full training set, including class-conditional information. Although the additional storage is small, the method does not eliminate the need to access the original dataset and may be unsuitable when data access or privacy constraints are central motivations for coreset selection. This weakens the claim that the method provides a broadly applicable solution to data reduction.

Questions for authors
  1. Since CCJ and QDA use statistics computed from the full training set, how sensitive are the results to the accuracy of these statistics, particularly when only a small subset of the data is available or class labels are noisy?
1 comment
AuthorsOpenReview ↗

Response to Reviewer dXeM

Thanks for taking the time to review our submission. It is true that as you pointed out, CCJ and QDA use the full dataset statistics when the coreset is being selected from the full dataset. Sensitivity and data privacy is an interesting challenge, but that is mostly a consideration in the related field of dataset distillation. In coreset selection we always provide a non-anonymised coreset by definition, and the only way to account for sensitivity of data would be to remove it from the training set before computing the coreset. Our answer to Q1 below experimentally tests this setting. We further added two experiments to loosen the reliance on full dataset statistics (App J).

Q1. How sensitive are the results to the stored statistics when only a subset of the data is available or the labels are noisy?

The QDA uses a mean and a covariance per class from the training set, so it is more sensitive than CCJ, particularly at the extremes. With ten images per class, it loses 2.67 points (Tab 35). From 250 images per class onwards the loss is only 0.34 points. The QDA is also robust to moderate label noise: replacing 10% or 20% of the labels costs only 0.15 and 0.35 points, and only heavy noise costs more, 1.37 points at 60% (Tab 36).

CCJ uses the pooled within-class variance of each DCT coefficient, from the dataset. We introduced a new experiment to test the effects of using the variance calculated from a subset on CCJ. CCJ's performance gain drops only by 0.09 when we calculate the variance from 10% of the CIFAR-100 dataset. If we use the coreset's own images, its gain decreases by 0.5. CCJ is robust to noisy labels, and we observe that the noisy labels actually provide a small improvement (Tab 34) but this is within run-to-run variance. We have added a more detailed discussion of this in App J.

Reviewer_3szaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies coreset training in the low-IPC (images per class) regime. The authors argue that, with very few training examples, models tend to learn features along high-variance directions even when low-variance features are more predictive of class labels. They investigate this behavior using a planted task that separates feature availability (how readily a model learns a feature) from predictivity (how informative that feature is about the label).

The paper proposes Coreset Covariance Jitter (CCJ), which uses covariance statistics estimated from the full training set to inject stronger noise along high-variance directions during student training. It further combines the student with a Quadratic Discriminant Analysis (QDA) classifier, built from full-dataset class statistics, to recover useful high-variance information at inference time. Experiments cover CIFAR-10, CIFAR-100, and ImageNet-1K; at IPC-10, CCJ+QDA improves all evaluated coreset selectors on CIFAR-100 and ImageNet-1K.

Critical strengths
  1. Compatibility and low storage overhead. CCJ can be applied after coreset selection without modifying the selector. Its fixed DCT/color basis and compact covariance statistics also limit additional storage requirements.
  2. Broad experimental coverage. The paper evaluates 16 selectors on CIFAR-100 and 11 on ImageNet-1K, alongside different IPC budgets, six student architectures, and a label-free setting. Separate CCJ and QDA results help identify their contributions.
  3. An interpretable analytical perspective. Distinguishing feature availability from predictivity provides a useful framework for investigating feature preference in low-data training.
Critical weaknesses
  1. Weak presentation and argumentative structure. The Abstract and Introduction state the central motivation too briefly, without sufficient supporting discussion or references. The connections among the theoretical analysis, practical method, and empirical results are not clearly developed.
  2. Insufficient justification for the central motivation. Claim 1 concerns the relative ease of learning high-variance features; it does not establish that relying on them harms generalization or that low-variance features are inherently preferable. Prior work, such as For Better or For Worse? Learning Minimum Variance Features With Label Augmentation, also shows that low-variance features can encode spurious correlations. The conditions under which suppressing high-variance information is beneficial remain unclear.
  3. A substantial gap between theoretical assumptions and practical settings. Appendix B explicitly acknowledges that natural images are not Gaussian, the small-signal condition is violated in the planted ConvNet experiments, and the kernel-diagonal assumption does not hold exactly for real networks. Moreover, the practical method replaces the analyzed eigenspaces with a fixed DCT/color basis and pooled within-class statistics, without sufficiently establishing that the relevant theoretical relationships are preserved.
  4. Incomplete mechanistic and design justification. The evidence does not convincingly isolate improved learning of useful low-variance features as the cause of CCJ's gains, as opposed to a more general regularization effect. The necessity of combining training-time suppression with a separate test-time QDA expert is also insufficiently established. Given existing covariance-based augmentation methods, the methodological novelty would benefit from sharper clarification.
Questions for authors
  1. Why should high-variance features be suppressed? Can the authors provide more direct evidence on natural datasets that reliance on high-variance directions limits generalization at low IPC? Under what conditions does CCJ help, especially when low-variance features contain spurious correlations?

  2. How does the theory justify CCJ in practice? Appendix B explicitly states that several theoretical assumptions are not satisfied by natural images, the planted ConvNet experiments, or real networks. Which conclusions can still be expected to hold in these settings, and on what basis? How is the theoretical availability argument preserved when moving to pooled within-class covariance statistics and a fixed DCT/color basis?

  3. What makes the CCJ–QDA division of labor necessary? I understand the intended rationale: CCJ reduces the student's reliance on high-variance cues so that it can learn low-variance features, while QDA separately retains high-variance class information. However, what evidence shows that separating these roles across training and inference is preferable to balancing both types of features within the student model? What distinctive benefit does this design provide beyond the individual improvements of CCJ and QDA, and how does it differ substantively from existing covariance-aware augmentation approaches?

  4. How can the near-additive pattern in Table 2 be explained? For all 11 selectors at IPC-10, the reported mean accuracies approximately satisfy:

    Acc(CCJ+Q)≈Acc(CCJ)+Acc(std+Q)−Acc(std).\mathrm{Acc}(\mathrm{CCJ+Q})\approx\mathrm{Acc}(\mathrm{CCJ})+\mathrm{Acc}(\mathrm{std+Q})-\mathrm{Acc}(\mathrm{std}).

    The mean absolute deviation from this relation is approximately 0.12 percentage points. Can the authors explain this consistent pattern and the full evaluation configurations needed to verify it?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3sza

Thanks for taking the time to review our submission. We added a paragraph clarifying the motivation in the introduction, and revised the text throughout.

Q1. Why reduce reliance on high-variance features, what direct evidence is there on natural data, and what if the low-variance features are spurious?

This is an important clarification. High-variance directions are not harmful at low IPC, but relying on them exclusively is. Based on Figure 11, we can see that this is true for natural data, as networks trained only with the high-variance features perform poorly when compared to training on the rest. CCJ helps by allowing the model to learn from the low-variance directions in addition to the high-variance ones. A new experiment shows that CCJ improves performance on unmodified & low-variance-only inputs (Tab 16) and the same on natural images (Tab 17).

Spurious correlations are possible, since CCJ ranks directions by the same within-class variance as Chidambaram & Ge (ICLR 2025). We didn't originally measure robustness to this as we used standard coreset selection evaluation protocols to provide consistency with prior work (i.e. the distributions of the train and test sets are matched). We now test CCJ under spurious correlations by planting a class-correlated pattern in CIFAR-100 (App M). CCJ makes the network follow a low-variance pattern at about half the strength the standard student needs (Tab 42), so if such a feature breaks at test time, CCJ makes the network more likely to rely on it. QDA reduces how often the CCJ student follows the low-variance pattern and reduces this effect. We agree that robustness to spurious correlations is important, and we hope that measuring this behaviour in our work will encourage future coreset selection techniques to also report it

Q2. Which theoretical conclusions still hold given the violated assumptions, the DCT basis & the pooled statistics?

We agree there is a gap between practical & theoretical settings. However, we note that the assumptions made are standard in the field (linearised, kernel-regime training: Jacot et al., 2018; Lee et al., 2019; Pezeshki et al., 2021; learning that decouples along the eigendirections of the kernel or of whitened data: Saxe et al., 2014; Bordelon et al., 2020; Simon et al., 2023; planted synthetic features: Hermann et al., 2024). We introduce new experiments to measure the deviation in t⋆ when we relax each of these assumptions (App K). In App L, we have added new results that compares CCJ in the DCT frame with CCJ in the eigenframe, including the measured t⋆ in the planted task (Tab 40). We measured the overlap between the top-K subspace of the basis using the DCT atoms and the top-K eigenvectors of the within-class covariance (Tab 41) and they overlap around 85% in the subspace spanned by the top 160 directions. Pooling the within-class statistics is exact in the planted task, because the backgrounds do not depend on the label. On CIFAR-100 we compared each class's covariance with the pooled one, and added this comparison to the revision (App L). At the median class, 94.0% of the pooled jitter's variance lies along a class's own 80 highest-variance directions. Therefore, pooling still targets the same directions as the ones we described in the theory. We hope this quantifies exactly how much these assumptions hold in practice

Q3. Why split work between CCJ & the QDA instead of balancing both feature types in the student. How does CCJ differ from covariance-based augmentation?

We have clarified that CCJ aims to balance both feature types by reducing over-reliance on dominant features. A CCJ student relies less on the high-variance subspace, and the QDA provides information that is not used by the network. Under our default protocol the two gains are close to additive at IPC-10 (see Q4). Under strong augmentation (Tab 15), a CCJ student gains clearly more from the QDA than one trained without CCJ, for every selector at IPC-10. App A.7.1 compares CCJ with Fancy PCA and ISDA, two augmentations that also use data statistics such as the covariance (Tabs 18 and 19).

Q4. Why do the CCJ & QDA gains add almost exactly in Tab 2. What configuration verifies it?

The paper notes this in App D.3, where we attribute it to CCJ & QDA preferring different subspaces. We have now made it clearer. QDA is a test-time approach. Therefore, when QDA is applied, the trained model weights (either std or +CCJ) are already fixed. Similarly, because QDA is driven by the statistics of the training set, it does not depend on the trained model. Since neither of their outputs affects the other, and assuming both approaches target non-redundant failure modes, additive performance is plausible. This is more visible for ImageNet-1K than CIFAR-100 simply because the gains (and thus variance) is lower. We have added the evaluation protocol for Tab 2 (App H), and are happy to provide code support reproducibility.

Dear reviewer, Thanks once again for your feedback! We put quite a lot of effort into creating an updated draft (which you should be able to view as a diff by downloading it). In particular, as detailed in our reply, we added a large number of additional experiments around the assumptions, which we believe has greatly solidified the foundation of the work. Please do let us know if you want us to paraphrase the findings here (we tried to keep our response above concise and within 1 form entry for your ease of reading) or if you have any further questions! Yours sincerely, The authors.

Reviewer_HBJPOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
1(−1)
Soundness
2
Read review
Paper summary
  • The draft presents an interesting problem: when we keep only a few images per class (i.e., the low IPC coreset setting) from a large dataset, neural networks tend to learn the wrong features.
  • Specifically, they learn features from directions where the data varies a lot (high-variance directions), even when features from low-variance directions would be more useful for classification.
  • The authors propose two things to fix this. First, Coreset Covariance Jitter (CCJ): a simple noise augmentation added during training. The noise is stronger in high-variance directions, making them less reliable and forcing the model to rely on low-variance directions instead.
  • Second, Quadratic Discriminant Analysis (QDA): a classifier added at test time to recover the class information that CCJ deliberately suppresses from the neural network. Together, CCJ+QDA improves accuracy over many coreset selection methods on CIFAR-10, CIFAR-100, and ImageNet-1K at IPC=10.
Critical strengths
  1. The paper clearly explains why the behavior they observe exists. The two claims (high-variance features are easier to learn; this imbalance shrinks as IPC increases) are stated precisely and tested carefully.

  2. Broad empirical validation. The method is tested on 18 coreset selectors for CIFAR-100 and 11 for ImageNet-1K. It works across different student architectures (ResNet-18, ResNet-50, VGG, DenseNet, ViT-Tiny). It also works without labels, which is a practical bonus.

  3. Very low cost. CCJ adds only 2-4% extra training time. The extra storage is tiny: about 0.34 images per class equivalent on CIFAR-100 and 0.01 on ImageNet-1K. This makes it very practical.

  4. Good ablation study. The paper tests many design choices: different noise shapes, different hyperparameter values, and comparisons with related augmentations (Fancy PCA, ISDA, whitening, RandAugment).

Critical weaknesses
  1. The formal proof of Claim 2 is missing for cross-entropy loss, which is proven only for squared loss. For cross-entropy, the authors rely only on empirical evidence. This is the loss that everyone actually uses in practice. This is a real gap between the theory and the method. The authors acknowledge this, but it is still a significant limitation.

  2. The assumptions may be strong. The theory requires the kernel to stay fixed during training, that backgrounds are Gaussian, and that the kernel is diagonal in the whitened basis. Practical neural networks trained with gradient descent may not satisfy these conditions.

  3. Performance sometimes drops at higher IPC. At IPC=50, CCJ+QDA hurts some selectors on ImageNet-1K. Given that the method is designed for low IPC, how does a practitioner know when not to use QDA?

  4. Comparison with stronger baselines is missing. The paper compares well against standard coreset methods and some augmentation techniques. However, it does not compare against methods that specifically address the feature learning problem, such as those that use pre-trained features (e.g., selecting coresets using DINO/CLIP embeddings and then training from scratch), or against more modern dataset distillation methods beyond the DC-BENCH set evaluated in the appendix. A direct comparison of CCJ+QDA against state-of-the-art dataset distillation methods under the same protocol would significantly strengthen the paper.

  5. The DCT basis approximation is not fully justified. CCJ uses a fixed DCT+chromatic basis instead of the true eigenbasis of the within-class covariance. The authors show this works empirically and argue it approximates natural image covariance well. But no direct comparison is shown between using the true covariance and the DCT approximation on the same dataset. This leaves it unclear how much accuracy is lost by the approximation.

  6. VGG-11 behavior is unexplained. The paper notes that VGG-11 behaves differently from other architectures and says it will be explored in future work. This is an honest admission, but it leaves a gap: does the theory actually apply to all architectures, or only to those without skip connections?

Questions for authors
  1. For cross-entropy loss, can you provide even a partial theoretical explanation (not a full proof) for why Claim 2 holds? The empirical evidence is convincing, but the theoretical gap is significant.

  2. Can you compare CCJ+QDA against coreset selection using pre-trained self-supervised features (e.g., using DINOv2 for selection and then CCJ during student training)?

  3. For the DCT approximation: can you measure the accuracy drop on CIFAR-100 IPC=10 when using the true within-class covariance versus the DCT basis? This would tell us how much the approximation costs.

  4. On ImageNet at IPC=50, what is the recommended practice (use CCJ but not QDA)? Can you give a simple rule for when to apply QDA?

  5. Can you clarify why DUAL performs so poorly on ImageNet (Table 2) but is one of the best methods on CIFAR-100? The appendix explains this, but the main paper does not.

3 comments
AuthorsOpenReview ↗

Response to Reviewer HBJP

Thanks for taking the time to review our submission. We first go through some of the weakness that you highlighted, and then we will address the questions. The theoretical proofs are derived under simplifying assumptions, and the experiments deviate from them. However, we would note that these assumptions are standard in the field (Jacot et al., 2018; Lee et al., 2019; Saxe et al., 2014; Bordelon et al., 2020; Pezeshki et al., 2021; Simon et al., 2023). We have now included in the paper a set of new experiments that try to identify the effect of each assumption and the degree to which they are violated. We measured the deviation of t⋆ when we move from a linearised kernel to the feature-learning case, and from real backgrounds to Gaussian backgrounds (App K.1, Tab 37). We have also measured the deviation from the diagonal case of the kernel to the non-diagonal case of the kernel (App K.2, Tab 38). Loosening these assumptions is one of our planned future works, but we hope that quantifying these effects will help the reader place the current work in its appropriate context.

In our analysis of the effects of CCJ with different architectures (App A.5), we show that CCJ improves every architecture, including VGG-11 (Tab 10). To find why VGG-11 improves the least, we have added new experiments that change one component of VGG-11 at a time. The main factor we found is how VGG-11 downsamples. Replacing its max-pooling layers with stride-2 convolutions, as in ResNet-18, raises its gain to +6.10. A more detailed analysis is added in App A.5.2 (Tab 13).

Q1. Can you give at least a partial theoretical explanation of Claim 2 under cross-entropy?

This was a challenging theoretical addition, but we agree that it significantly strengthens the plausibility of the work. We have added a proof for Claim 2 under cross-entropy loss. It is presented in App G, requiring only one further assumption.

Q2. How does CCJ+QDA do on coresets selected with self-supervised features such as DINOv2?

This is an interesting experiment which we hadn't previously considered. We have now expanded the main papers results by additionally testing two new methods of coreset selection that use DINOv2 features. We have added ProbCover (Yehuda et al., 2022) and herding with DINOv2 features to Tab 1. We observe that our method improves the performance in this setting as well.

Q3. How much accuracy does the DCT basis cost against the true within-class eigenbasis on CIFAR-100 at IPC-10?

It is an important point that we never quantified the cost of this modification. We have now evaluated the performance with the true within-class eigenbasis in CIFAR-100 and in the planted task, and the values are provided in Tabs 39 and 40. From Tab 39, we observe that using the true eigenvectors, the performance improves by +0.32 points on average. However, we note that this gain in performance is quite minor compared to the significantly increased storage cost, and therefore the DCT basis still seems like the optimal compromise.

Q4. On ImageNet-1K at IPC-50, should one use CCJ without the QDA, and is there a simple rule for when to apply the QDA?

On ImageNet-1K at IPC-50, we recommend using CCJ without the QDA. QDA works really well at low IPC and with a low number of classes. The performance degrades a lot when we increase the number of classes, as shown in the case of ImageNet. We have expanded the discussion in the paper to make these rules of thumb more explicit to the reader. We also note that QDA is a test-time classifier, and therefore it is very cheap to validate whether it improves the performance in any given situation.

Q5. Why does DUAL perform poorly on ImageNet-1K (Tab 2) but well on CIFAR-100?

The implementation of DUAL in Tab 2 uses the same class-balanced top-K sampler as all other compared techniques, rather than their proposed β-sampler. This was done to ensure consistency of comparison in the main paper, while the "native" analysis of DUAL plus β-sampler was relegated to the Appendix. In retrospect, this may have shown DUAL in an unfavourable light, and we have now reversed their positions (native β-sampler in main paper, and match top-k sampler in the appendix). We hope this will avoid confusion for future readers. We weren't able to run other baselines like GraNd and DeepFool because their computational cost was too high.

Reviewer_HBJPOpenReview ↗

Thanks for the rebuttal

I thank the authors for providing their rebuttal. After carefully reviewing all the review comments, the rebuttal, and the submission, I believe it is not ready for acceptance in its current form.

My primary concerns are two:

i) the difficulty in reading and understanding the paper. The paper took me an unusually long time to understand. This is also noted by another reviewer (3sza). I strongly believe the paper needs a careful rewrite to better assimilate the claims, analyses, and the proofs.

ii) justification/motivation for the proposed method. The relationship between high-variance (or low-variance) features and their generalizability needs to be discussed convincingly. In its current form, there is insufficient discussion to appreciate under which settings suppressing which features (low- or high-variance) is meaningful.

Further, the validity of the assumptions needs to be properly presented.

Dear reviewer, Thanks once again for your feedback! We apologise if this paper took a lot of effort to follow, in the updated draft we have made significant attempts to smooth the flow of the story for the reader. If you have any further suggestions for areas that could be clarified please do let us know!

Regarding the justification/motiviation of the proposed method, we would kindly like to draw your attention to our response to your co-reviewer which we believe has largely addressed this issue. We tried to keep our direct response concise for ease of reading, but the full details should be visible in the updated paper pdf.

\textbf{Q1. Why reduce reliance on high-variance features, what direct evidence is there on natural data, and what if the low-variance features are spurious?} This is an important clarification. High-variance directions are not harmful at low IPC, but relying on them exclusively is. Based on Figure 11, we can see that this is true for natural data, as networks trained only with the high-variance features perform poorly when compared to training on the rest. CCJ helps by allowing the model to learn from the low-variance directions in addition to the high-variance ones. A new experiment shows that CCJ improves performance on unmodified & low-variance-only inputs (Tab 16) and the same on natural images (Tab 17). Spurious correlations are possible, since CCJ ranks directions by the same within-class variance as Chidambaram & Ge (ICLR 2025). We didn't originally measure robustness to this as we used standard coreset selection evaluation protocols to provide consistency with prior work (i.e. the distributions of the train and test sets are matched). We now test CCJ under spurious correlations by planting a class-correlated pattern in CIFAR-100 (App M). CCJ makes the network follow a low-variance pattern at about half the strength the standard student needs (Tab 42), so if such a feature breaks at test time, CCJ makes the network more likely to rely on it. QDA reduces how often the CCJ student follows the low-variance pattern and reduces this effect. We agree that robustness to spurious correlations is important, and we hope that measuring this behaviour in our work will encourage future coreset selection techniques to also report it

\textbf{Q3. Why split work between CCJ & the QDA instead of balancing both feature types in the student. How does CCJ differ from covariance-based augmentation?} We have clarified that CCJ aims to balance both feature types by reducing over-reliance on dominant features. A CCJ student relies less on the high-variance subspace, and the QDA provides information that is not used by the network. Under our default protocol the two gains are close to additive at IPC-10 (see Q4). Under strong augmentation (Tab 15), a CCJ student gains clearly more from the QDA than one trained without CCJ, for every selector at IPC-10. App A.7.1 compares CCJ with Fancy PCA and ISDA, two augmentations that also use data statistics such as the covariance (Tabs 18 and 19).

Our feeling is that with the above changes, the justification for the approach in the current draft is quite clearly stated, but we would definitely appreciate your advice on how this can be improved further!

Please do let us know if you have any follow-on questions we can answer. Yours sincerely, The authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.