acceptodds
Under review as a conference paper at ICLR 2027

Enhancing few-shot samples for Deep Imbalanced Regression via Frobenius norm

Abstract

Deep Imbalanced Regression (DIR) aims to learn a regression model where the training label distribution is extremely imbalanced. Recent studies in DIR have shown that aligning feature ordinality with that of labels can effectively enhance the prediction outcomes. However, these gains are largely confined to many-shot samples, while medium- and few-shot regions benefit little. This suggests that preserving feature-label ordinality alignment alone may be insufficient. Empirically, we further observe that feature intensity, measured by the Frobenius norm, also exhibits a clear ordinal structure primarily shaped by well-fitted many-shot samples. Specifically, samples whose feature intensities deviate from this structure consistently exhibit higher test errors, whereas those that better conform to it tend to achieve lower errors. Therefore, this motivates us to investigate the feature intensity to further improve the model performance. In this paper, we propose Frobenius-norm Enhanced Inference (FEI), a plug-and-play post-training pipeline that requires only a three-layer MLP to distill the feature-norm prior from a pretrained model and exploits the learned ordinal property to improve performance on medium- and few-shot samples. Furthermore, we theoretically prove that this constraint leads to a tighter generalization bound under mild assumptions in DIR. Extensive experiments on real-world datasets show that FEI improves medium- and few-shot performance by over 4% and 6%, respectively, at the cost of less than 2% on many-shot samples, which further demonstrates the effectiveness of our method.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Deep Imbalanced Regression, Frobenius norm

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_miadOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces Frobenius-norm Enhanced Inference (FEI) for Deep Imbalanced Regression (DIR). Their method builds upon the empirical observation that in target regions with plenty of data, the features exhibit an approximately ordinal pattern in the their Frobenius norms and their error is low. In contrast, the feature Frobenius norms of medium- and few-shot samples do not follow this ordinal patter and have a higher error. They conclude that a label-dependent feature magnitude is required for every target label, meaning that small targets should correspond to smaller Frobenius norms of their latent features as bigger targets. This motivates the authors to propose a two-step approach:

  1. they obtain the Frobenius norm of the model latent features z of every many-shot sample. Then they train a 3-layer MLP with the datasets original target label as an input and the Frobenius norm as the "new" target label
  2. They take this MLP and predict the Frobenius norm of samples of medium and few-shot regions. Based on the output of the MLP they rescale the feature representation and retrain the regressor on these rescaled features They test their method on AgeDB-DIR and IMDB-WIKI-DIR as also on WIND in the appendix.
Critical strengths
  • The paper links feature norms, known to matter for regression under distribution shift (DAR), to DIR. The empirical observation that feature norms follow an ordinal pattern in the many-shot region and deviate in the tail, where errors are high, is a useful diagnostic.
  • It uses three datasets and eight recent baselines, averaged over five seeds.
Critical weaknesses
  • The most critical weakness is about the method itself. The retraining of r~\tilde{r} is done on prototypes of the rescaled latent features z^\hat{z}, but the rescaling mechanism doesn't apply at test time. The 3-layer MLP gg takes the ground-truth labels as input, so test features can never be rescaled and the regressor r~\tilde{r} gets the unscaled zz as input. The gains most plausibly come from retraining the head on the balanced per-label prototypes and not from the fact that they are rescaled. An ablation that retrains r~\tilde{r} on unscaled prototypes is needed to show that the rescaling contributes anything.
  • Fig. 5 contradicts Table 1 and the text: according to Section 2.3, test features are never rescaled, so "MAE after rescale" on test data shouldn't exist. If "MAE after rescale" means the comparison between the regressor rr and r~\tilde{r}, why does Figure 5 show an improvement in every single bin for all four baselines, while Table 1 reports worse many-shot MAE in every case? Also, the "MAE after rescale" for RankSim is so small in the many-shot bins that it is not readable anymore, which does not add up with the numbers in Table 1.
  • Some information is missing for reproducibility and traceability:
    • No information about the backbone architecture used (based on previous literature I would guess ResNet-50, but there is no confirmation).
    • No information about batch sizes, the training epochs of r~\tilde{r}, or whether r~\tilde{r} is trained from scratch or initialized from rr.
    • They have tested across 5 seeds, but only report the mean.
    • Appendix G has an empty link, and I cannot find any supplementary material or code as claimed in the "Reproducibility Statement".
  • The UVote MAE values (for both UVote and UVote+Ours) appear to be swapped between Tables 1 and 2, while the GM values seem to be in the correct tables.
  • Issues regarding presentation:
    • Parenthetical citations use \cite{} / \citet{} instead of \citep{} throughout the paper.
    • The use of the word "distill" in Section 2.3 is misleading, since there is no teacher–student setup or anything similar.
    • Fig. 3 uses the same color for two legend items.
Questions for authors
  1. How does performance change when r~\tilde{r} is retrained on unscaled per-label prototypes, i.e., without the MLP gg? How do you explain the gains, given that test features are never rescaled?
  2. How exactly was "MAE after rescale" in Fig. 5 computed? Were test features rescaled using ground-truth test labels?
  3. Can you report the standard deviations across the five seeds?
4 comments

Inference mismatch The reason we do not use rescale at inference is : Rescaling is a label-conditioned operation used during head retraining. Its coefficient, (\alpha(z,y)=g(y)/|z|_2), requires the ground-truth label, which is unavailable at inference. FEI therefore uses rescaled prototypes to adjust the regression head during post-training, and predicts with the resulting head on original features, (\hat y=\tilde r(f(x))). The intended benefit is carried by the adjusted head rather than by a test-time transformation. We acknowledge that the resulting training–inference mismatch is a limitation to assess, rather than assuming it is harmless.Our decision to freeze the encoder is motivated by decoupled learning. [3] show that representations learned from imbalanced data can remain effective, while adjusting the prediction head improves long-tailed recognition. [4] extend this principle to imbalanced regression through regressor retraining with a frozen encoder, e.g., RRT.FEI follows this principle and introduces label-conditioned magnitude calibration to construct the prototypes used for head retraining.The objective is to retain the learned representation while adjusting how the regressor maps features to target values. Since the rescaling coefficient requires the ground-truth label, calibration is performed during post-training, and inference uses the original features with the adjusted regressor. Our D-versus-FEI ablation evaluates this design directly: both use unscaled test features, yet FEI improves medium-shot MAE and few-shot MAE over most baselines.

Problem on Fig. 5 Sorry for not making this mis-understanding, the left legend in Fig.5 is actually the training set per-label MAE. We aim to show that FEI can help to construct the ordinal relationship by rescaling on the training features across different methods. We will make it clear in the revised version.

Experiment setup The result presented in the main paper is based on the pre-trianed weights from the previous works (for encoders). Backbone architecture: AgeDB-DIR and IMDB-WIKI-DIR (ResNet50), WIND-DIR (ResNet-18) Batch sizes: train from scratch, MLP (bs:256), regressor (bs:8) Training epochs of regressor: train from scratch, MLP (300 iterations), regressor (50 iterations)

UVote results Thanks for pointing it out, the result should be and we will revise it in the revised version. AgeDB-DIR

MethodMAE overallMany.Med.Few.G-Mean overallMany.Med.Few.
UVote6.816.387.588.664.273.974.865.86
UVote+Ours6.806.417.498.594.314.054.775.82

IMDB-WIKI-DIR

MethodMAE overallMany.Med.Few.G-Mean overallMany.Med.Few.
UVote7.346.7712.0121.224.023.826.2710.71
UVote+Ours7.266.7611.3220.603.963.805.6211.49

The use of "distill" Thank you for requesting clarification. We use “distill” in the broader sense of extracting and compactly encoding knowledge, rather than referring to a conventional teacher–student framework. This broader usage has precedent: Dataset Distillation [1,2] compresses knowledge from a training dataset into a small set of synthetic examples. Reference In FEI, the pretrained encoder provides label–feature-norm observations from many-shot regions, and the lightweight network (g) summarizes this information into a label-conditioned magnitude prior for underrepresented regions. Thus, “distill” describes the extraction of this specific prior into a compact predictor. We will explicitly define this intended meaning in Section 2.3 to distinguish it from conventional teacher–student knowledge distillation.

Fig.3 Thanks for pointing it out, we have removed the "Balanced exposure" legend in our revised version. We aimed to show the dataset we used is balanced but it seems redundant.

Ablation study and standard deviation

To separate calibration from prototype averaging and label balancing, we will conduct the following controlled ablations:

VariantFeature rescalingHead-training inputs
A (orignal model)NoNo
BNoIndividual features
CYesIndividual features
DNoPer-label prototypes
OursYes
MethodVariantMAE-AllManyMedFew
VANILLAA7.76286.73339.432912.8664
VANILLAB7.6986 ± 0.00106.7913 ± 0.00109.0619 ± 0.007712.4971 ± 0.0036
VANILLAC7.6940 ± 0.00376.8022 ± 0.00569.0215 ± 0.034012.4451 ± 0.0078
VANILLAD7.6939 ± 0.00006.8255 ± 0.00008.9181 ± 0.000012.5105 ± 0.0000
VANILLAOurs7.6591 ± 0.00296.9303 ± 0.05498.6470 ± 0.142211.8110 ± 0.1631
BalancedMSEA7.50666.61819.213511.1770
BalancedMSEB7.5650 ± 0.01246.5505 ± 0.00719.5651 ± 0.072611.6161 ± 0.0307
BalancedMSEC7.4976 ± 0.00056.6370 ± 0.00179.1284 ± 0.007811.1157 ± 0.0017
BalancedMSED7.5810 ± 0.00256.5344 ± 0.00169.6308 ± 0.014511.7970 ± 0.0078
BalancedMSEOurs7.4944 ± 0.00446.7079 ± 0.00948.9265 ± 0.017810.9625 ± 0.0193
RankSimA6.91416.33827.79539.9167
RankSimB6.9604 ± 0.03986.3135 ± 0.07998.0038 ± 0.338410.1842 ± 0.3435
RankSimC6.9111 ± 0.02836.3912 ± 0.07687.7164 ± 0.27049.5935 ± 0.2880
RankSimD6.9153 ± 0.00376.3259 ± 0.00637.8188 ± 0.02859.9831 ± 0.0294
RankSimOurs6.9105 ± 0.03806.4006 ± 0.06587.7080 ± 0.29409.5207 ± 0.3092
ConRA7.42716.50798.752112.4447
ConRB7.6628 ± 0.40166.8095 ± 0.18209.1711 ± 1.080711.5503 ± 0.6174
ConRC7.4766 ± 0.06896.4910 ± 0.03039.0610 ± 0.321412.4033 ± 0.2723
ConRD8.0679 ± 0.28157.2040 ± 0.07579.7922 ± 0.934711.4589 ± 0.4362
ConROurs7.4009 ± 0.00986.5605 ± 0.00428.6357 ± 0.048011.9236 ± 0.0400
OrdinalEntropyA7.59026.58869.157212.7162
OrdinalEntropyB7.5432 ± 0.00256.6298 ± 0.00098.9214 ± 0.016812.3586 ± 0.0060
OrdinalEntropyC7.5403 ± 0.00196.6329 ± 0.00108.9055 ± 0.012112.3349 ± 0.0039
OrdinalEntropyD7.5406 ± 0.00006.6327 ± 0.00008.9083 ± 0.000012.3329 ± 0.0000
OrdinalEntropyOurs7.4688 ± 0.03996.8471 ± 0.01858.2150 ± 0.138011.2773 ± 0.0297
SRLA7.22526.63748.28499.8466
SRLB7.2560 ± 0.01556.5379 ± 0.00118.5643 ± 0.059410.4195 ± 0.0506
SRLC7.2091 ± 0.03366.6448 ± 0.02128.1963 ± 0.18229.8086 ± 0.1395
SRLD7.1622 ± 0.00076.6011 ± 0.00148.0409 ± 0.006110.0317 ± 0.0058
SRLOurs7.1937 ± 0.03426.7206 ± 0.12147.9409 ± 0.19189.5952 ± 0.1739

[1] Dataset distillation. arXiv preprint arXiv:1811.10959 (2018).

[2] Dataset Distillation for Pre-Trained Self-Supervised Vision Models. NeurIPS 2025

[3] Decoupling Representation and Classifier for Long-Tailed Recognition. ICLR 2020.

[4] Delving into Deep Imbalanced Regression. ICML 2021.

Thanks for the reply, I will update my score to weak rejection. The most critical weakness is answered but the overall paper had too many presentation issues, missing informations and ablations that I cannot justify a higher increase

Thanks for your kind comments; we have uploaded a revised version of our paper. In our paper, we 1) changed the legend and added an explanation to Fig. 5; 2) removed the redundant legend in Fig.3; 3) added the backbone architecture, batch size, and experimental details; 4)we also fixed the empty link; 5) revised the UVote MAE values in Tab.1,2; 6) changed every \cite{} to \citep{}; 6) added a set of demonstrations on explaining the recalled feature; 7) added ablation studies in the Appendix; 8) fixed the presentation issues.

We hope our revised version addresses your concerns. And thanks again for your comments.

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

This paper addresses deep imbalanced regression (DIR) by proposing Frobenius-norm Enhanced Inference (FEI), a lightweight and plug-and-play post-training module. The proposed method leverages the feature Frobenius norm and trains an MLP to learn the relationship between target labels and feature norms from well-represented many-shot regions. The learned relationship is then used to rescale features in medium- and few-shot regions, followed by prototype-based regressor retraining. Experiments on three real-world datasets demonstrate improvements in underrepresented regions across several existing DIR methods.

Critical strengths
  1. An interesting research perspective. The paper provides a valuable perspective on deep imbalanced regression by investigating feature magnitudes through the Frobenius norm, rather than focusing solely on feature-label ordinality.
  2. Lightweight and flexible design. The proposed FEI module is computationally efficient and can be integrated into various existing DIR methods without retraining the entire backbone, making it potentially useful in practical applications.
Critical weaknesses
  1. Insufficient justification for the transferability of the learned feature-norm relationship. The paper does not provide sufficient theoretical or empirical evidence to establish that the feature-norm relationship learned from many-shot regions can be reliably transferred to medium- and few-shot regions. Although the authors demonstrate an approximately ordinal relationship between labels and feature norms in well-represented regions, it remains unclear whether this relationship generalizes to underrepresented regions. Furthermore, the experimental results show considerable variations in performance improvements across datasets involving similar tasks, such as AgeDB-DIR and IMDB-WIKI-DIR, while some configurations even exhibit performance degradation. These observations raise concerns about the robustness and generalizability of the proposed approach.
  2. Insufficient ablation studies. The paper lacks comprehensive ablation experiments to evaluate the contribution of individual components. Since FEI involves both feature-norm rescaling and prototype-based regressor retraining, it is difficult to determine how much of the performance improvement is directly attributable to the proposed Frobenius-norm calibration. Additional experiments investigating alternative feature norms and network architectures would also help demonstrate the effectiveness of the current design choices.
Questions for authors
  1. Ablation studies and architectural choices. Could the authors provide more detailed ablation studies to isolate the contributions of feature-norm rescaling and prototype-based regressor retraining? Additionally, have the authors investigated whether alternative feature norms or more expressive network architectures for learning the label-to-norm mapping could lead to different results?
  2. Consistency between post-training and inference. During post-training, feature rescaling is applied to construct the feature prototypes used for regressor retraining. However, according to Section 2.3, the inference pipeline directly uses the original feature extractor and the retrained regressor without feature rescaling. Could the authors clarify the rationale behind this design choice? Does the difference in feature distributions between post-training and inference introduce any potential performance degradation?
4 comments

Relationship on generalizing to underrepresented regions We directly examine whether additional supervision in many-shot regions benefits few-shot prediction by varying the number of many-shot training samples while keeping the few-shot training data and evaluation set fixed. Increasing the many-shot sample count from 300 to 1,000 per label reduces few-shot MAE from 27.429 to 26.255 for Vanilla and from 25.203 to 22.930 for FEI, corresponding to reductions of 4.3% and 9.0%, respectively. Vanilla exhibits intermediate fluctuations, whereas FEI improves monotonically over these controlled sample-count settings. These results provide empirical evidence that additional many-shot supervision can benefit underrepresented regions despite their unchanged training data. Furthermore, FEI outperforms Vanilla in every evaluated setting, with relative few-shot MAE reductions of 7.1%–13.4%, including the original training distribution. Together with the rescaled-versus-unscaled prototype ablation, these findings support both cross-region learning and the additional benefit of FEI’s norm-calibration mechanism.

Many-shot sample numbers (Per-label)Few-shot MAE (Ours)Few-shot MAE (Vanilla)
30025.20327.429
50023.08826.241
70023.07326.641
100022.93026.255
Orignal22.87524.632

configurations even exhibit performance degradation. Thank you for this comment. The primary objective of FEI is to improve prediction in underrepresented, particularly few-shot regions. Improving the few-shot performance is crucial in many real-world senarios, espically in rare diease detection [1]. Therefore, our goal is to imrpove the few-shot sample performance instead of only concentrating on the overall performance which normally improved by the many-shot performance. Overall error aggregates performance across label regions and can obscure these targeted improvements. For example, FEI reduces Vanilla’s few-shot MAE from 13.67 to 11.81 on AgeDB-DIR and from 25.89 to 25.07 on IMDB-WIKI-DIR. Although the improvement magnitudes differ, both results support the intended benefit to few-shot samples. The trade-off is particularly clear on WIND-DIR: for ConR, few-shot MAE decreases from 14.19 to 11.77, while overall MAE increases from 7.62 to 7.81. Evaluating this configuration solely by overall MAE would overlook the substantial improvement in the underrepresented region. Therefore, FEI should be assessed primarily through few-shot performance together with the accompanying many-shot and overall trade-offs. We acknowledge that some configurations and metrics deteriorate the many-shot regions but our claim concerns improved performance in underrepresented regions, rather than uniform improvement across every metric.

Although AgeDB-DIR and IMDB-WIKI-DIR both involve age estimation, they differ substantially in dataset size, label distribution, and data composition. For example, IMDB-WIKI-DIR contains approximately 191.5K training images, compared with 12.2K for AgeDB-DIR. Sharing the same prediction task therefore does not imply identical performance gains. More importantly, FEI primarily aims to improve underrepresented, particularly few-shot, regions. For Vanilla, few-shot MAE decreases from 13.67 to 11.81 on AgeDB-DIR and from 25.89 to 25.07 on IMDB-WIKI-DIR. Despite different improvement magnitudes, both results support this objective. Overall MAE can obscure such benefits. On WIND-DIR, for example, ConR’s few-shot MAE improves from 14.19 to 11.77, while overall MAE increases from 7.62 to 7.81. This reflects a trade-off between label regions. Accordingly, our claim is that FEI can improve underrepresented-region performance across different datasets, with the accompanying many-shot and overall trade-offs explicitly reported.

Alternative feature norms We also report the few-shot performance obetween different feature norm, we show that our method on the feature Frobenius norm can provide the most significant performance improvement compared to other norm methods, which also consistent with the previous works [4,5] that leverages the Frobenius norm in practice.

Norm MethodVanillaRankSimCon-RSRL
Frobenius norm (Ours)12.886 ± 1.79011.246 ± 1.45211.202 ± 0.43412.636± 0.603
Min-Max Norm13.289 + 1.27513.033 ± 1.38014.348 ± 0.44313.082± 0.490
Standardization13.817+ 0.43712.689 ± 2.07015.092 ± 0.88912.807± 0.422

Post-training and inference mismatch The reason we do not use rescale at inference is : Rescaling is a label-conditioned operation used during head retraining. Its coefficient, (\alpha(z,y)=g(y)/|z|_2), requires the ground-truth label, which is unavailable at inference. FEI therefore uses rescaled prototypes to adjust the regression head during post-training, and predicts with the resulting head on original features, (\hat y=\tilde r(f(x))). The intended benefit is carried by the adjusted head rather than by a test-time transformation. We acknowledge that the resulting training–inference mismatch is a limitation to assess, rather than assuming it is harmless.Our decision to freeze the encoder is motivated by decoupled learning. [3] show that representations learned from imbalanced data can remain effective, while adjusting the prediction head improves long-tailed recognition. [2] extend this principle to imbalanced regression through regressor retraining with a frozen encoder, e.g., RRT.FEI follows this principle and introduces label-conditioned magnitude calibration to construct the prototypes used for head retraining.The objective is to retain the learned representation while adjusting how the regressor maps features to target values. Since the rescaling coefficient requires the ground-truth label, calibration is performed during post-training, and inference uses the original features with the adjusted regressor. Our D-versus-FEI ablation evaluates this design directly: both use unscaled test features, yet FEI improves medium-shot MAE and few-shot MAE over most baselines.

Ablation study on rescaled and unscaled features:

VariantFeature rescalingHead-training inputs
A (orignal model)NoNo
BNoIndividual features
CYesIndividual features
DNoPer-label prototypes
OursYes
MethodVariantMAE-AllManyMedFew
VANILLAA7.76286.73339.432912.8664
VANILLAB7.6986 ± 0.00106.7913 ± 0.00109.0619 ± 0.007712.4971 ± 0.0036
VANILLAC7.6940 ± 0.00376.8022 ± 0.00569.0215 ± 0.034012.4451 ± 0.0078
VANILLAD7.6939 ± 0.00006.8255 ± 0.00008.9181 ± 0.000012.5105 ± 0.0000
VANILLAOurs7.6591 ± 0.00296.9303 ± 0.05498.6470 ± 0.142211.8110 ± 0.1631
BalancedMSEA7.50666.61819.213511.1770
BalancedMSEB7.5650 ± 0.01246.5505 ± 0.00719.5651 ± 0.072611.6161 ± 0.0307
BalancedMSEC7.4976 ± 0.00056.6370 ± 0.00179.1284 ± 0.007811.1157 ± 0.0017
BalancedMSED7.5810 ± 0.00256.5344 ± 0.00169.6308 ± 0.014511.7970 ± 0.0078
BalancedMSEOurs7.4944 ± 0.00446.7079 ± 0.00948.9265 ± 0.017810.9625 ± 0.0193
RankSimA6.91416.33827.79539.9167
RankSimB6.9604 ± 0.03986.3135 ± 0.07998.0038 ± 0.338410.1842 ± 0.3435
RankSimC6.9111 ± 0.02836.3912 ± 0.07687.7164 ± 0.27049.5935 ± 0.2880
RankSimD6.9153 ± 0.00376.3259 ± 0.00637.8188 ± 0.02859.9831 ± 0.0294
RankSimOurs6.9105 ± 0.03806.4006 ± 0.06587.7080 ± 0.29409.5207 ± 0.3092
ConRA7.42716.50798.752112.4447
ConRB7.6628 ± 0.40166.8095 ± 0.18209.1711 ± 1.080711.5503 ± 0.6174
ConRC7.4766 ± 0.06896.4910 ± 0.03039.0610 ± 0.321412.4033 ± 0.2723
ConRD8.0679 ± 0.28157.2040 ± 0.07579.7922 ± 0.934711.4589 ± 0.4362
ConROurs7.4009 ± 0.00986.5605 ± 0.00428.6357 ± 0.048011.9236 ± 0.0400
OrdinalEntropyA7.59026.58869.157212.7162
OrdinalEntropyB7.5432 ± 0.00256.6298 ± 0.00098.9214 ± 0.016812.3586 ± 0.0060
OrdinalEntropyC7.5403 ± 0.00196.6329 ± 0.00108.9055 ± 0.012112.3349 ± 0.0039
OrdinalEntropyD7.5406 ± 0.00006.6327 ± 0.00008.9083 ± 0.000012.3329 ± 0.0000
OrdinalEntropyOurs7.4688 ± 0.03996.8471 ± 0.01858.2150 ± 0.138011.2773 ± 0.0297
SRLA7.22526.63748.28499.8466
SRLB7.2560 ± 0.01556.5379 ± 0.00118.5643 ± 0.059410.4195 ± 0.0506
SRLC7.2091 ± 0.03366.6448 ± 0.02128.1963 ± 0.18229.8086 ± 0.1395
SRLD7.1622 ± 0.00076.6011 ± 0.00148.0409 ± 0.006110.0317 ± 0.0058
SRLOurs7.1937 ± 0.03426.7206 ± 0.12147.9409 ± 0.19189.5952 ± 0.1739

[1] Dist Loss: Enhancing Regression in Few-Shot Region through Distribution Distance Constraint. ICLR 2025.

[2] Delving into Deep Imbalanced Regression. ICML 2021.

[3] Decoupling Representation and Classifier for Long-Tailed Recognition. ICLR 2020.

[4] Representation subspace distance for domain adaptation regression. ICML 2021.

[5] DARE-GRAM : Unsupervised Domain Adaptation Regression by Aligning. CVPR 2023.

Reviewer_DjreOpenReview ↗

The central soundness concern regarding cross-region transferability remains only partially resolved

I think the authors have satisfactorily addressed Question 1 and Question 2, but the central soundness concern regarding cross-region transferability remains only partially resolved. My primary concern remains the core assumption underlying FEI, namely the label-to-feature-norm relationship learned from many-shot regions can be reliably transferred to medium- and few-shot regions.

While the new experiments show that increasing many-shot supervision improves few-shot performance and that FEI consistently outperforms the corresponding vanilla baselines, they do not directly validate the transferability of the learned norm-label mapping itself. This does not fully address my concern. Therefore, I prefer to keep the score at 2.

Thank you for acknowledging that Questions 1 and 2 have been satisfactorily addressed.

Thank you for clarifying your concern. The rationale for transferring information across label regions stems from the continuous structure of regression: different label regions represent related portions of the same prediction task. Under the assumption that label-conditioned feature statistics vary smoothly, well-represented regions can provide useful structural information for underrepresented ones. Therefore, the motivation for cross-region information sharing has precedent in DIR. [1] explained that a sparsely sampled target benefits from a well-represented neighborhood and introduced FDS to transfer feature statistics between nearby label bins. Similarly, [2] borrowed information from samples with similar labels to improve representations for underrepresented targets. [3] even brought the similarities between groups of data to enhance the underrepresented samples. These works support exploiting relationships across continuous labels, although they do not establish the transferability of FEI’s specific norm mapping.

FEI applies this principle specifically to feature magnitude. Many-shot samples provide better-supported observations for learning a label-conditioned norm relationship. Crucially, FEI does not assign the same many-shot norm to all underrepresented labels; the learned mapping conditions the calibration magnitude on each target label. This prior guides prototype construction and regressor retraining while keeping the encoder fixed. The relevant assumption is therefore that the label-conditioned norm relationship retains useful structure across regions, rather than that many-shot and few-shot features have identical distributions. Label continuity motivates this assumption but does not guarantee it, particularly when underrepresented labels lie far outside the many-shot range. We position the transferred relationship as a structured inductive prior for learning the regressor, rather than a universally exact model of feature norms.

[1] Delving into Deep Imbalanced Regression. ICML 2021.

[2] Variational Imbalanced Regression: Fair Uncertainty Quantification via Probabilistic Smoothing. NeurIPS 2023.

[3] Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression. AAAI 2025.

Reviewer_sBK8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Frobenius-norm Enhanced Inference (FEI), a lightweight plug-and-play method for improving pretrained models for deep imbalanced regression. Motivated by an observed relationship between target labels and feature norms in many-shot regions, FEI trains a small MLP to predict feature norms from labels, then uses the predictions to rescale medium-shot / few-shot features, retrains the regression head using per-label feature prototypes. Experiments on AgeDB-DIR, IMDB-WIKI-DIR and WIND-DIR show improvements in medium and few-shot regions over baselines, with a small tradeoff in many-shot regions.

Critical strengths
  • The authors present an intuitive observation that the feature norms of labels exhibit an ordinal relationship with target labels in many-shot regions. The method inspired by this relationship is also simple and well-motivated.

  • The proposed method can be applied to different baseline models on the DIR task with small computational overhead.

  • The experimental demonstrations are extensive with multiple datasets and baselines.

Critical weaknesses
  1. My major concern is the technical novelty. As the authors already discussed in the manuscript, the idea of feature Frobnenius norm is built on [1, 2] which also target at regression problems, and SRL [3] also uses the feature norm directly on DIR problems. The authors should discuss more, even consider the experimental comparison against all these three work in the main manuscript.

  2. The contribution of feature-norm rescaling still needs more ablations to demonstrate its effetiveness. The proposed method retrains the regressor with one prototype per label in phase 2. This is similar to regressor re-training in [4] which is modified from decoupling training of imbalanced classification. I encourage the authors to discuss the relationship and differences between the methods. Also, some more ablations: 1. retrain the regressor on prototypes without rescaling with the same training settings; and 2. rescale the features and retrain the regressor on individual data points, instead of one prototype per label. These ablations can help make the paper more solid.

[1] Chen et al., Representation subspace distance for domain adaptation regression. ICML 2021

[2] Adachi et al., Test-time adaptation for regression by subspace alignment. ICLR 2025.

[3] Dong et al., Improve Representation for Imbalanced Regression through Geometric Constraints, CVPR 2025.

[4] Yang et al., Delving into Deep Imbalanced Regression. ICML 2021

Questions for authors
  • Could the authors provide experimental comparisons using prototypes without rescaling under the same training settings?

  • I'm concerned about the Few. MAE for UVote in Table 2. The value of 8.66 and 8.59 are much lower than other baselines, and their corresponding GM are higher than MAE values. This is impossible as GM is a geometric mean of absolute errors and GM <= MAE holds for the same subset. Please double check these numbers and clarify how the numbers are calculated.

4 comments

technical concern

We present the different between our work and [1,2,3] both from problem setup, solution, and experiment setup as follows:

MethodSetupUnlabeled target inputs used for adaptation?Main mechanism
[1] Chen et al.: RSDUnsupervised domain adaptation regressionYes, during adaptation trainingAlign source–target representation subspaces while preserving feature scale
[2] Adachi et al.: SSATest-time adaptation for regressionYes, during test-time adaptationAlign feature distributions within an output-relevant subspace
[3] Dong et al.: SRLDeep imbalanced regressionNoModel feature-manifold geometry using normalized representations and geometric constraints
Ours: FEIDeep imbalanced regression with post-training head adjustmentNoLearn a label-conditioned magnitude prior from many-shot training regions to calibrate underrepresented prototypes

difference between the decouped learning We agree that regressor retraining itself is not our novelty. FEI follows the decoupling principle of freezing the learned representation and retraining the prediction head; we do not claim head retraining itself as novel. Its distinguishing mechanism is label-conditioned feature-magnitude calibration: FEI learns a norm prior from many-shot regions and uses it to construct calibrated prototypes for underrepresented labels. Standard head retraining rebalances the contribution of existing features, whereas FEI additionally modifies their training-time magnitudes before fitting the head.

*prototypes w and w/o rescaling * To separate calibration from prototype averaging and label balancing, we will conduct the following controlled ablations:

VariantFeature rescalingHead-training inputs
A (orignal model)NoNo
BNoIndividual features
CYesIndividual features
DNoPer-label prototypes
OursYes
MethodVariantMAE-AllManyMedFew
VANILLAA7.76286.73339.432912.8664
VANILLAB7.6986 ± 0.00106.7913 ± 0.00109.0619 ± 0.007712.4971 ± 0.0036
VANILLAC7.6940 ± 0.00376.8022 ± 0.00569.0215 ± 0.034012.4451 ± 0.0078
VANILLAD7.6939 ± 0.00006.8255 ± 0.00008.9181 ± 0.000012.5105 ± 0.0000
VANILLAOurs7.6591 ± 0.00296.9303 ± 0.05498.6470 ± 0.142211.8110 ± 0.1631
BalancedMSEA7.50666.61819.213511.1770
BalancedMSEB7.5650 ± 0.01246.5505 ± 0.00719.5651 ± 0.072611.6161 ± 0.0307
BalancedMSEC7.4976 ± 0.00056.6370 ± 0.00179.1284 ± 0.007811.1157 ± 0.0017
BalancedMSED7.5810 ± 0.00256.5344 ± 0.00169.6308 ± 0.014511.7970 ± 0.0078
BalancedMSEOurs7.4944 ± 0.00446.7079 ± 0.00948.9265 ± 0.017810.9625 ± 0.0193
RankSimA6.91416.33827.79539.9167
RankSimB6.9604 ± 0.03986.3135 ± 0.07998.0038 ± 0.338410.1842 ± 0.3435
RankSimC6.9111 ± 0.02836.3912 ± 0.07687.7164 ± 0.27049.5935 ± 0.2880
RankSimD6.9153 ± 0.00376.3259 ± 0.00637.8188 ± 0.02859.9831 ± 0.0294
RankSimOurs6.9105 ± 0.03806.4006 ± 0.06587.7080 ± 0.29409.5207 ± 0.3092
ConRA7.42716.50798.752112.4447
ConRB7.6628 ± 0.40166.8095 ± 0.18209.1711 ± 1.080711.5503 ± 0.6174
ConRC7.4766 ± 0.06896.4910 ± 0.03039.0610 ± 0.321412.4033 ± 0.2723
ConRD8.0679 ± 0.28157.2040 ± 0.07579.7922 ± 0.934711.4589 ± 0.4362
ConROurs7.4009 ± 0.00986.5605 ± 0.00428.6357 ± 0.048011.9236 ± 0.0400
OrdinalEntropyA7.59026.58869.157212.7162
OrdinalEntropyB7.5432 ± 0.00256.6298 ± 0.00098.9214 ± 0.016812.3586 ± 0.0060
OrdinalEntropyC7.5403 ± 0.00196.6329 ± 0.00108.9055 ± 0.012112.3349 ± 0.0039
OrdinalEntropyD7.5406 ± 0.00006.6327 ± 0.00008.9083 ± 0.000012.3329 ± 0.0000
OrdinalEntropyOurs7.4688 ± 0.03996.8471 ± 0.01858.2150 ± 0.138011.2773 ± 0.0297
SRLA7.22526.63748.28499.8466
SRLB7.2560 ± 0.01556.5379 ± 0.00118.5643 ± 0.059410.4195 ± 0.0506
SRLC7.2091 ± 0.03366.6448 ± 0.02128.1963 ± 0.18229.8086 ± 0.1395
SRLD7.1622 ± 0.00076.6011 ± 0.00148.0409 ± 0.006110.0317 ± 0.0058
SRLOurs7.1937 ± 0.03426.7206 ± 0.12147.9409 ± 0.19189.5952 ± 0.1739

Problem on UVote

Thank you for catching this error. You are correct that GM cannot exceed MAE when both are computed from the same absolute errors on the same subset. The UVote and UVote+FEI few-shot MAE values in Table 2 were incorrectly transcribed when preparing the table for AgeDB-DIR and IMDB-WIKI_DIR. We apologize for this reporting error. The corrected entries for AgeDB-DIR and IMDB-WIKI_DIR are: AgeDB-DIR

MethodMAE overallMany.Med.Few.G-Mean overallMany.Med.Few.
UVote6.816.387.588.664.273.974.865.86
UVote+Ours6.806.417.498.594.314.054.775.82

IMDB-WIKI-DIR

MethodMAE overallMany.Med.Few.G-Mean overallMany.Med.Few.
UVote7.346.7712.0121.224.023.826.2710.71
UVote+Ours7.266.7611.3220.603.963.805.6211.49

We will correct this mistake in our main paper. Thanks again for pointing it out!

I thank the authors for the rebuttal and the additional ablations. The B/C/D variants are helpful, but they also suggest that the contribution of feature-norm rescaling is limited and inconsistent. For ConR, retraining on unscaled prototypes (D) results in a lower few-shot MAE than FEI. As the authors also agree that regressor retraining is not novel, my concern on technical novelty stays same. Therefore I keep my original score.

Thank you for examining the additional ablations. We would like to clarify both the contribution of rescaling and the distinction from conventional regressor retraining.

The controlled comparison between D and FEI isolates the contribution of rescaling: both retrain the regressor using per-label prototypes, while FEI additionally calibrates its feature norms. Across the six base methods, this calibration reduces medium-shot MAE in all six cases and few-shot MAE in five. For example, few-shot MAE decreases from 12.5105 to 11.8110 for Vanilla and from 12.3329 to 11.2773 for OrdinalEntropy. We acknowledge the ConR exception: D achieves a lower few-shot MAE than FEI (11.4589 vs. 11.9236, 4% lower). However, FEI significantly improves its medium-shot MAE (9.7922 to 8.6357, 11% improvement) and overall MAE (8.0679 to 7.4009, 7.5% improvement). Thus, the results support a broadly beneficial, but not universal, contribution from rescaling.

Our claimed novelty is the transfer of a learned label-conditioned feature-norm prior, rather than regressor retraining itself. Instead of the contribution on training paradigm (e.g., decoupled training), our contribution is that FEI learns this relationship from well-represented many-shot regions and uses it to calibrate underrepresented prototypes that guide regressor learning. Retraining serves as the mechanism for incorporating this prior into the predictor. Variant D already includes prototype-based retraining but lacks this transfer mechanism; the additional improvements with FEI therefore provide evidence that its benefits extend beyond retraining alone.

We hope this distinction and the controlled comparisons help clarify the scope of our contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.