acceptodds
Withdrawn from ICLR 2027

Cross-Scale Consistency on Common Nyquist Support

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-scale consistency is invalid outside the information jointly representable by both views: downsampling removes frequencies above the low-resolution Nyquist limit and may alias them into misleading components. We introduce common-support consistency, which projects both representations onto their shared spectral support and attenuates pairs with high alias risk; the gate follows the sampling relation rather than learned semantic confidence. On CIFAR-100, NYQUIST-VALID CONSISTENCY reaches 77.5% accuracy at 32 pixels and 64.1% at a 16-pixel stress resolution, with negative-transfer area 0.1 and unsupported consistency gradient ratio (UCGR) 0.09. Fixed low-pass consistency reaches 76.6%, 61.8%, 0.4, and 0.22, while alias weighting without the projector reaches 76.9%, 62.3%, 0.2, and 0.18. The controls isolate common-support projection from generic smoothing or loss reweighting, establishing a sampling-theoretic validity principle for cross-scale representation learning at a measured training throughput of 1438 images/s versus 1715 without consistency.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
cross-scale representation learning, consistency regularization, multi-scale learning, Nyquist sampling, spectral representation, aliasing, resolution robustness, frequency-domain learning

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

This research addresses representation learning for images at varying resolutions. Since downsampling often leads to a loss of information, enforcing global consistency across different resolutions is undesirable. To overcome this, we propose a method that constrains the model to maintain consistency only for the frequency components shared between the original and downsampled images. In our experiments, we applied this approach to the CIFAR-100 image classification task and observed improvements in accuracy at lower resolutions.

Critical strengths

The idea of ​​enforcing consistency in representations across images of different resolutions is interesting.

It is also commendable that the authors conducted experiments on the CIFAR-100 dataset using two different resolutions: 32x32 and 16x16, to demonstrate the effectiveness of their proposed method.

Critical weaknesses

It is unclear that the reason and background for achieving consistent representations across varying resolutions are not sufficiently explained.

The review of related work is not sufficient. In the introduction, the authors state that imposing consistency constraints on the entire feature representation is undesirable; however, they fail to cite any relevant prior work, leaving the motivation of this problem formulation unclear.

The descriptions of the methodology are significantly lacking, making it impossible to reproduce the results. Furthermore, because the experimental settings are not clearly defined, it is difficult to assess the validity of the reported results.

The experiments are limited to the CIFAR-100 dataset; given that the method is intended for image classification, it should be evaluated on a dataset such as ImageNet (with 224px resolution) at least.

Questions for authors

Please describe the proposed method more precisely. Please show the other experimental results to show the effectiveness of the proposed method.

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

This paper investigates the potential negative transfer caused by enforcing feature consistency across different image resolutions. To address this issue, the authors propose Nyquist-Valid Consistency (NVC), which restricts cross-scale consistency to the common spectral support of different views and introduces an alias-risk-based reliability weight. The paper also presents the Unsupported Consistency Gradient Ratio (UCGR) to evaluate the proposed mechanism. Experiments on CIFAR-100 report improved classification accuracy and reduced negative transfer compared with the evaluated baselines, at the cost of additional training computation.

Critical strengths
  1. The paper identifies a meaningful problem in cross-scale representation learning, where enforcing full feature consistency across different resolutions may introduce unreliable constraints due to information loss and aliasing.
Critical weaknesses
  1. The paper does not specify how the spectral projector is constructed, how alias risk and reliability weights are calculated, or how the proposed loss is integrated into the overall training objective. Furthermore, the theoretical connection between input-level Nyquist constraints and learned feature representations is not adequately justified, making the method difficult to reproduce and assess.
  2. Experiments are limited to CIFAR-100 at two resolutions (16px and 32px), without reporting statistical variability across multiple random seeds or comparisons with competitive existing methods. The baselines primarily focus on component-level ablations, while key metrics such as UCGR and negative-transfer area lack precise definitions.
  3. The paper is too short and the references are too few.
Questions for authors
  1. Could the authors provide the exact definitions and implementations of the spectral projector, alias-risk function, reliability weighting mechanism, and complete training objective?
  2. What theoretical assumptions justify applying input-level Nyquist constraints to learned feature representations? Under what conditions does Eq. (3) hold, and how is UCGR calculated?
  3. Could the authors clarify how negative-transfer area is computed and report results across multiple random seeds? Additionally, how does the proposed method compare with existing cross-resolution methods across more datasets and resolutions?
Reviewer_zyYvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Nyquist-valid consistency for cross-scale representation learning. It restricts consistency to shared frequency support and uses alias-risk weighting to reduce unreliable supervision. Experiments on CIFAR-100 show improved accuracy and reduced negative transfer under resolution changes.

Critical strengths

The common-support formulation provides a reasonable way to avoid imposing consistency on frequency components lost during downsampling.

Critical weaknesses
  1. The textual content is very poor, with too little content to present many innovations and details of the paper!
  2. The experimental evaluation is restricted to CIFAR-100. A single dataset is insufficient to support a general claim about cross-scale representation learning.
  3. Only one main resolution-shift setting is evaluated. It is unclear whether the method works under larger scale changes, non-integer scaling factors, or different resizing operators.
  4. The common-support projector is not sufficiently specified. The exact spectral mask, cutoff frequencies, and projection layer are missing, making the method difficult to reproduce.
  5. The alias-risk function and its weighting rule are not explicitly defined. These are central components of the proposed objective rather than minor implementation details.
  6. The method applies sampling-theoretic arguments to learned feature maps without establishing that their frequency components correspond to the observable frequencies of the input images.
  7. The theoretical analysis is limited to a local gradient bound. Equation (3) does not establish that minimizing the proposed objective improves generalization or prevents negative transfer.
Questions for authors
  1. What are the exact definitions of the spectral projector and alias-risk weighting function?
  2. How is the input-image Nyquist limit mapped to the spectral support of nonlinear learned features?
  3. Can the authors provide results across multiple datasets, resizing operators, and random seeds?
  4. How are UCGR and negative-transfer area computed?
  5. Please also see Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.