Evaluating Rank-N-Contrast: Continuous and Robust Representations for Regression

Alexandre Chidiac; Arkin Worlikar; Valentin Six

REVIEW 4 major objections 5 minor 9 references

Evaluating Rank-N-Contrast: Continuous and Robust Representations for Regression

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This replication claims that adding the Rank-N-Contrast loss at the encoder improves regression accuracy and robustness to missing target ranges, provided the dataset is large enough.

desk verdict A well-intentioned replication of Rank-N-Contrast that never actually measures the thing it claims (regression accuracy), so the central conclusion is unsupported as written. read the letter →

arxiv 2411.16298 v3 pith:G6H3FFWV submitted 2024-11-25 cs.LG stat.ML

Valentin Six , Alexandre Chidiac , Arkin Worlikar This is my paper · ORCID

classification cs.LGstat.ML

keywords Rank-N-Contrastregressioncontrastivelearningrepresentationlossfunctionageestimationrobustnessmissingdata

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 replication study asks whether the Rank-N-Contrast loss, which orders an encoder's embeddings by the rank of the target value, benefits regression models in practice. The authors compare a plain L1 pipeline with a pipeline whose encoder is first trained with $L_{RNC}$ and then combined with L1 prediction, using the AgeDB face-age dataset. They report that the RNC-trained model converges faster and reaches roughly half the L1-only training loss, and that it degrades less when ages 30--40 are removed from the training set. Extending the test to a 552-image mango-weight dataset, the advantage reverses: L1 converges faster and the final $L_{RNC}$ value is higher. The paper concludes that, given enough data, using $L_{RNC}$ at the encoder improves regression performance and robustness to missing ranges in the training set.

What carries the argument

The load-bearing object is the Rank-N-Contrast loss $L_{RNC} = \frac{1}{2N}\sum_{i=1}^{2N} \ell_{RNC}(i)$, where each per-anchor term contrasts the anchor's embedding with other samples in the batch weighted by their rank in label space: samples with closer target values contribute larger similarity terms in a temperature-scaled softmax. This loss is what forces the encoder to place samples with nearby continuous labels nearby in the embedding, giving the later L1 predictor a representation that respects label order. The paper's experiments vary only this loss at the encoder while keeping the same predictor loss.

What would settle it

Run the same AgeDB split with identical hyperparameters, optimizer, and seeds for L1 and RNC+L1; if the RNC model's validation loss on the full age range, including ages 30--40, is not below the L1 model's, or if the gap reverses under re-seeding, the paper's central robustness claim fails.

Watch

Extended reading notes

Core claim

The claim the authors are trying to establish is that a regression-aware contrastive loss at the encoder produces a latent space ordered by the target value, and that this ordering carries over into better and more robust predictions. On AgeDB, $L_{RNC}+L1$ reaches a final training loss near 1 while L1 converges near 2, and its validation curve starts low and stays smoother. With ages 30--40 missing from training, both models' validation losses rise from 6.75 to 8.5, but the L1 model's convergence becomes notably noisier and slower, which the authors read as evidence that the continuous embedding helps extrapolate into the missing range. On MangoMassNet-552, the pattern breaks: L1 converges faster on both training and validation, and the final encoder $L_{RNC}$ loss is 5.37 versus 3.92--3.98 on AgeDB. The stated conclusion is therefore conditional: with enough data, $L_{RNC}$ at the encoder helps regression and robustness.

Load-bearing premise

The paper assumes that the L1 baseline and the RNC+L1 model are trained in exactly the same way except for the added representation loss, but it gives no hyperparameters, seeds, or training budget to back that up.

Editorial extensions

If this is right

With enough training data, adding $L_{RNC}$ to the encoder lowers final regression loss and speeds convergence relative to L1-only training.
The continuous ordering learned by $L_{RNC}$ makes validation performance degrade more gracefully when a contiguous range of target values is absent from training.
The benefit is not automatic on small datasets: on 552 images, L1 converges faster, so RNC's usefulness appears to scale with data or label diversity.
RNC is a plug-in loss for the encoder, so it can be attached to existing regression predictors without changing the prediction layer.

Reading between the lines

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

The paper leaves implicit that batch composition controls when RNC helps: if a batch contains few distinct target values, the ranking signal weakens, which may explain why the 552-image mango dataset loses the advantage.
The missing-age-range result suggests a potential application to regression settings with censored or sparse labels, such as medical or pricing data, but the paper tests only age on one dataset, so transfer remains an open question.
A natural next experiment is to vary AgeDB size continuously, for example from 500 to 16,000 images, to map where the RNC advantage over L1 appears or disappears.

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.

Desk Editor's Note

read the letter

Quick take: this is a student replication of Rank-N-Contrast (RNC) with a small robustness twist, and the write-up is honest about being a replication. But the central claim — that RNC improves regression performance and robustness — is never actually measured. What you get instead are training and validation loss curves for the RNC loss itself, plus a table of final RNC loss values. Lower RNC loss at the encoder does not equal lower MAE or RMSE on the regression task, and the abstract's "state-of-the-art performance" is asserted without any SOTA comparison.

What's genuinely there: the authors clearly describe their setup, reuse the original RNC code as they say, and add two reasonable experiments: a holdout of ages 30–40 on AgeDB and a smaller mango-weight dataset. The idea of testing robustness to missing target ranges is sensible, and the AgeDB loss curves do show RNC converging faster and to a lower loss than plain L1. That part is a fair, if shallow, reproduction.

The soft spots are proportionate to how soft they actually are. No final MAE/RMSE anywhere, so the main empirical benefit is inferred, not shown. No seeds, no error bars, no hyperparameter reporting (learning rate, batch size, epochs, optimizer), so you cannot rule out confounding between the RNC loss and other training choices. The mango experiment—where L1 converges faster than RNC+L1 on both train and validation—is a direct counterexample to the generality claim, and the authors hand-wave it with "the learner at the encoder is not learning as much," without quantifying prediction error. That is a gap in the evidence, not a theoretical contradiction, but it is a load-bearing gap.

Who gets value: this is useful as a cautionary example for a reading group on replication standards, or as a starting point for someone who wants to replicate RNC and needs a minimal baseline. It is not a contribution that changes anyone's assessment of RNC.

My recommendation for peer review: I would not send this to referees in its current form. The claim is measurable and the authors just haven't measured it. A resubmission with a proper metrics table (MAE/RMSE on both datasets), multiple seeds, and a toned-down abstract would be worth another look, but as it stands the manuscript doesn't support its own conclusion.

Referee Report

4 major / 5 minor

Summary. This manuscript is a replication/extension study of the Rank-N-Contrast (RNC) method for regression. The authors compare a model trained with standard L1 loss against a model that first trains the encoder with the RNC loss and then fine-tunes with L1, across three settings: the AgeDB dataset, an AgeDB variant with ages 30-40 removed from training, and the smaller MangoMassNet-552 dataset. The paper concludes that RNC yields better regression performance and robustness to missing data. The evidence presented is limited to training/validation loss curves (Figures 1-6) and the final RNC loss at the encoder (Table I); no held-out task metric such as MAE or RMSE is reported.

Significance. If substantiated, the manuscript would provide a useful independent check on the original RNC claim and extend it to a new small dataset and a missing-range generalization test. Strengths include the choice of an additional real-world regression dataset (mango mass estimation), a deliberately challenging holdout split, and transparency about reusing the original implementation (Section IV-A). However, the current evidence does not establish the claimed benefit: no regression accuracy metric is reported, no seeds or hyperparameters are given, and the mango experiment visually contradicts the benefit claim. The contribution is therefore presently a description of an experimental setup plus qualitative loss-curve observations, rather than a validated replication result.

major comments (4)

[Section VI / Table I] The central claim that RNC improves regression performance and robustness is not measured by any task metric. The only quantitative result reported is the final RNC loss at the encoder (Table I), and the supporting evidence is qualitative inspection of training/validation loss curves (Figures 1-6). A lower auxiliary representation loss or a lower L1 validation loss during training does not establish lower MAE or RMSE on a held-out test set, particularly because the prediction heads and training schedules differ between the two pipelines. Please report final test-set MAE/RMSE (with standard deviations over multiple seeds) for L1 and RNC+L1 on AgeDB, modified AgeDB, and MangoMassNet.
[Section IV-A and Section VI] The manuscript does not report any hyperparameters, optimizer settings, learning rates, batch sizes, epoch counts, seed values, or compute budgets, so the reader cannot verify that the L1 and RNC+L1 pipelines are otherwise identical. The two pipelines also differ by design (RNC+L1 first trains the encoder with L_RNC, then trains the whole model with L1, whereas L1 is trained end-to-end), so attribution of performance differences to L_RNC requires demonstrating that the comparison is controlled. Please include a complete experimental setup table and, ideally, release the code and seeds.
[Section VI, Figures 5-6] The mango dataset experiment is a counterexample to the paper's benefit claim: the L1 model converges faster than RNC+L1 on both training and validation loss. The explanation offered ("the learner at the encoder is not learning as much") is speculative, and no final prediction error is reported for either model. This result, as presented, undermines the general claim that RNC helps regression tasks and must be addressed with a task-metric comparison on this dataset rather than a post-hoc verbal explanation.
[Abstract and Section IV] The abstract claims the study "achieve[s] state-of-the-art performance," but no comparison is made to any state-of-the-art method beyond the L1 baseline; SupCon, DINO, and SimCLR are discussed in Section II but never evaluated. Please either remove the state-of-the-art claim or add the corresponding baselines.

minor comments (5)

[Section VI, AgeDB paragraph] The sentence "We obtain two trained regression models that we can for which we can compare the performances" is ungrammatical and should be rewritten.
[Section VI, MangoMassNet paragraph] The phrase "which significantly less than the previous dataset" is missing a verb; it should be "which is significantly less than...".
[Figures 1-6] All figures should include axis labels, legends, and captions specifying exactly which loss is plotted (e.g., L1 prediction loss or total loss), so that the curves are interpretable without relying on the prose.
[Table I] Table I reports only the final RNC loss at the encoder; including the final validation L1 loss for both L1 and RNC+L1 models would make the comparison more informative and would partially address the missing quantification in the text.
[Section V] This section lists potential challenges but the paper never returns to them in the results or discussion; either connect these items to the experimental findings or remove the section to avoid a dangling outline.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the paper is an under-supported replication, not a result reduced to its own inputs.

full rationale

This manuscript is a replication study of Rank-N-Contrast. It does not derive a result from its inputs; it reuses the original implementation from [8] and reports learning curves. There is no fitted parameter that is later renamed as a prediction, no equation in this paper that defines its conclusion into existence, and no self-citation chain: the present authors are not authors of [8]. The reuse of the original code and the lack of task metrics/seeds/hyperparameters make the empirical conclusion weak and potentially confounded, because the L1 and RNC+L1 pipelines differ in training schedule as described in Section VI, but those are evidence-quality problems, not circularity. The central claim is an empirical claim about regression performance that could in principle be falsified by an external benchmark; it is under-supported but not equivalent to the inputs by construction. Per the hard rules, missing support belongs to correctness risk, not to circularity. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim relies on the RNC loss formula from prior work, on the choice of hyperparameters inherited from the original code, and on treating validation loss curves as sufficient evidence of performance.

free parameters (4)

temperature tau in RNC loss
Inherited from original RNC code, not reported; affects contrastive scaling but is not fit in this paper.
embedding dimension de
Encoder output dimension, inherited from original code; not reported or varied.
training hyperparameters (learning rate, batch size, epochs)
Not reported; chosen by hand or inherited from original code; all results depend on them.
dataset split seed
Random splits for AgeDB and MangoMassNet are not described; single run per configuration.

assumptions (3)

domain assumption RNC loss and its theoretical guarantee are taken from Zha et al. [8] without derivation.
Section III-B reproduces the likelihood and loss formula but offers no proof; the paper's central claim rests on this external theory.
ad hoc to paper Validation loss is a reliable proxy for regression performance.
Section VI evaluates only loss curves and never reports test MAE or accuracy; the conclusion that RNC 'achieves better performance' depends on this proxy.
domain assumption The AgeDB and MangoMassNet labels are accurate and the mango images contain sufficient signal for weight estimation.
The paper relies on datasets from [5] and [3] without analyzing label noise or image quality.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Evaluating Rank-N-Contrast: Continuous and Robust Representations for Regression." pith.science (2026). https://pith.science/paper/G6H3FFWV

@misc{pith2026241116298,
  author       = {Pith},
  title        = {Pith review of: Evaluating Rank-N-Contrast: Continuous and Robust Representations for Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G6H3FFWV}},
  note         = {Machine review of arXiv:2411.16298}
}

read the original abstract

This document is an evaluation of the original "Rank-N-Contrast" (arXiv:2210.01189v2) paper published in 2023. This evaluation is done for academic purposes. Deep regression models often fail to capture the continuous nature of sample orders, creating fragmented representations and suboptimal performance. To address this, we reproduced the Rank-N-Contrast (RNC) framework, which learns continuous representations by contrasting samples by their rankings in the target space. Our study validates RNC's theoretical and empirical benefits, including improved performance and robustness. We extended the evaluation to an additional regression dataset and conducted robustness tests using a holdout method, where a specific range of continuous data was excluded from the training set. This approach assessed the model's ability to generalize to unseen data and achieve state-of-the-art performance. This replication study validates the original findings and broadens the understanding of RNC's applicability and robustness.

Figures

Figures reproduced from arXiv: 2411.16298 by the authors.

**Figure 2.** Validation Loss for L1 and LRNC + L1 on AgeDB dataset using [5], by removing all images in the training set that correspond to an age ranging from 30 to 40. The validation set contains images that correspond to the complete range of ages, which should be a challenge for the standard L1 model. By conducting this experiment, we wish to see if the continuity in the embedding space allows the model to perform better on … view at source ↗

**Figure 4.** shows that we have a similar convergence process, although the final loss value is higher than with the regular split (from 6.75 to 8.5) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (3 more)

**Figure 3.** Figure 3: Training Loss for L1 and LRNC + L1 on modified AgeDB dataset [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]

**Figure 5.** Figure 5: Training Loss for L1 and LRNC +L1 on MangoMassNet-552 dataset [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Validation Loss for L1 and LRNC + L1 on MangoMassNet-552 dataset Figures 5 and 6 clearly show us that both models converge faster, but the main difference with using a smaller dataset is that the L1 model converges faster than LRNC + L1, for both training and validatio…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 5 canonical work pages

[1]

͂ G1j=c& G 2Hغu+6oތW^9DDDDDDDD #

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

arXiv
[2]

Emerging properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650--9660, 2021

2021
[3]

A simple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597--1607. PMLR, 2020

2020
[4]

Estimating mango mass from rgb image with convolutional neural network

Mohammad Hafiz Bin Ismail, Mohd Nazuan Wagimin, and Tajul Rosli Razak. Estimating mango mass from rgb image with convolutional neural network. In 2022 3rd International Conference on Artificial Intelligence and Data Sciences (AiDAS) , pages 105--110. IEEE, 2022

work page 2022
[5]

Supervised contrastive learning

Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems , 33:18661--18673, 2020

2020
[6]

Agedb: the first manually collected, in-the-wild age database

Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshop , volume 2, page 5, 2017

work page 2017
[7]

Ordinal regression with multiple output cnn for age estimation

Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4920--4928, 2016

work page 2016
[8]

Dex: Deep expectation of apparent age from a single image

Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops , pages 10--15, 2015

work page 2015

Show all 9 references

[9]

Rank-n-contrast: learning continuous representations for regression

Kaiwen Zha, Peng Cao, Jeany Son, Yuzhe Yang, and Dina Katabi. Rank-n-contrast: learning continuous representations for regression. Advances in Neural Information Processing Systems , 36, 2024

2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

͂ G1j=c& G 2Hغu+6oތW^9DDDDDDDD #

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

arXiv

[2] [2]

Emerging properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650--9660, 2021

2021

[3] [3]

A simple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597--1607. PMLR, 2020

2020

[4] [4]

Estimating mango mass from rgb image with convolutional neural network

Mohammad Hafiz Bin Ismail, Mohd Nazuan Wagimin, and Tajul Rosli Razak. Estimating mango mass from rgb image with convolutional neural network. In 2022 3rd International Conference on Artificial Intelligence and Data Sciences (AiDAS) , pages 105--110. IEEE, 2022

work page 2022

[5] [5]

Supervised contrastive learning

Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems , 33:18661--18673, 2020

2020

[6] [6]

Agedb: the first manually collected, in-the-wild age database

Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshop , volume 2, page 5, 2017

work page 2017

[7] [7]

Ordinal regression with multiple output cnn for age estimation

Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4920--4928, 2016

work page 2016

[8] [8]

Dex: Deep expectation of apparent age from a single image

Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops , pages 10--15, 2015

work page 2015