- 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:
- 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
- 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 is done on prototypes of the rescaled latent features , but the rescaling mechanism doesn't apply at test time. The 3-layer MLP takes the ground-truth labels as input, so test features can never be rescaled and the regressor gets the unscaled 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 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 and , 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 , or whether is trained from scratch or initialized from .
- 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
- How does performance change when is retrained on unscaled per-label prototypes, i.e., without the MLP ? How do you explain the gains, given that test features are never rescaled?
- How exactly was "MAE after rescale" in Fig. 5 computed? Were test features rescaled using ground-truth test labels?
- Can you report the standard deviations across the five seeds?
4 comments
Ablation study and standard deviation
To separate calibration from prototype averaging and label balancing, we will conduct the following controlled ablations:
| Variant | Feature rescaling | Head-training inputs |
|---|---|---|
| A (orignal model) | No | No |
| B | No | Individual features |
| C | Yes | Individual features |
| D | No | Per-label prototypes |
| Ours | Yes |
| Method | Variant | MAE-All | Many | Med | Few |
|---|---|---|---|---|---|
| VANILLA | A | 7.7628 | 6.7333 | 9.4329 | 12.8664 |
| VANILLA | B | 7.6986 ± 0.0010 | 6.7913 ± 0.0010 | 9.0619 ± 0.0077 | 12.4971 ± 0.0036 |
| VANILLA | C | 7.6940 ± 0.0037 | 6.8022 ± 0.0056 | 9.0215 ± 0.0340 | 12.4451 ± 0.0078 |
| VANILLA | D | 7.6939 ± 0.0000 | 6.8255 ± 0.0000 | 8.9181 ± 0.0000 | 12.5105 ± 0.0000 |
| VANILLA | Ours | 7.6591 ± 0.0029 | 6.9303 ± 0.0549 | 8.6470 ± 0.1422 | 11.8110 ± 0.1631 |
| BalancedMSE | A | 7.5066 | 6.6181 | 9.2135 | 11.1770 |
| BalancedMSE | B | 7.5650 ± 0.0124 | 6.5505 ± 0.0071 | 9.5651 ± 0.0726 | 11.6161 ± 0.0307 |
| BalancedMSE | C | 7.4976 ± 0.0005 | 6.6370 ± 0.0017 | 9.1284 ± 0.0078 | 11.1157 ± 0.0017 |
| BalancedMSE | D | 7.5810 ± 0.0025 | 6.5344 ± 0.0016 | 9.6308 ± 0.0145 | 11.7970 ± 0.0078 |
| BalancedMSE | Ours | 7.4944 ± 0.0044 | 6.7079 ± 0.0094 | 8.9265 ± 0.0178 | 10.9625 ± 0.0193 |
| RankSim | A | 6.9141 | 6.3382 | 7.7953 | 9.9167 |
| RankSim | B | 6.9604 ± 0.0398 | 6.3135 ± 0.0799 | 8.0038 ± 0.3384 | 10.1842 ± 0.3435 |
| RankSim | C | 6.9111 ± 0.0283 | 6.3912 ± 0.0768 | 7.7164 ± 0.2704 | 9.5935 ± 0.2880 |
| RankSim | D | 6.9153 ± 0.0037 | 6.3259 ± 0.0063 | 7.8188 ± 0.0285 | 9.9831 ± 0.0294 |
| RankSim | Ours | 6.9105 ± 0.0380 | 6.4006 ± 0.0658 | 7.7080 ± 0.2940 | 9.5207 ± 0.3092 |
| ConR | A | 7.4271 | 6.5079 | 8.7521 | 12.4447 |
| ConR | B | 7.6628 ± 0.4016 | 6.8095 ± 0.1820 | 9.1711 ± 1.0807 | 11.5503 ± 0.6174 |
| ConR | C | 7.4766 ± 0.0689 | 6.4910 ± 0.0303 | 9.0610 ± 0.3214 | 12.4033 ± 0.2723 |
| ConR | D | 8.0679 ± 0.2815 | 7.2040 ± 0.0757 | 9.7922 ± 0.9347 | 11.4589 ± 0.4362 |
| ConR | Ours | 7.4009 ± 0.0098 | 6.5605 ± 0.0042 | 8.6357 ± 0.0480 | 11.9236 ± 0.0400 |
| OrdinalEntropy | A | 7.5902 | 6.5886 | 9.1572 | 12.7162 |
| OrdinalEntropy | B | 7.5432 ± 0.0025 | 6.6298 ± 0.0009 | 8.9214 ± 0.0168 | 12.3586 ± 0.0060 |
| OrdinalEntropy | C | 7.5403 ± 0.0019 | 6.6329 ± 0.0010 | 8.9055 ± 0.0121 | 12.3349 ± 0.0039 |
| OrdinalEntropy | D | 7.5406 ± 0.0000 | 6.6327 ± 0.0000 | 8.9083 ± 0.0000 | 12.3329 ± 0.0000 |
| OrdinalEntropy | Ours | 7.4688 ± 0.0399 | 6.8471 ± 0.0185 | 8.2150 ± 0.1380 | 11.2773 ± 0.0297 |
| SRL | A | 7.2252 | 6.6374 | 8.2849 | 9.8466 |
| SRL | B | 7.2560 ± 0.0155 | 6.5379 ± 0.0011 | 8.5643 ± 0.0594 | 10.4195 ± 0.0506 |
| SRL | C | 7.2091 ± 0.0336 | 6.6448 ± 0.0212 | 8.1963 ± 0.1822 | 9.8086 ± 0.1395 |
| SRL | D | 7.1622 ± 0.0007 | 6.6011 ± 0.0014 | 8.0409 ± 0.0061 | 10.0317 ± 0.0058 |
| SRL | Ours | 7.1937 ± 0.0342 | 6.7206 ± 0.1214 | 7.9409 ± 0.1918 | 9.5952 ± 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.
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
IMDB-WIKI-DIR
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.