Pith. sign in

REVIEW 3 major objections 4 minor 56 references

Prior2Posterior: Model Prior Correction for Long-Tailed Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that the right prior for long-tail correction is the one the model itself learned from its average predictions, and that reweighting by that prior provably balances the test class proportions and outperforms…

desk verdict Useful empirical recipe wrapped in an overstated theory: the marginal-matching theorem does not survive softmax normalization. read the letter →

arxiv 2412.16540 v1 pith:BCQXY6KH submitted 2024-12-21 cs.CV cs.AI

S Divakar Bhat , Amit More , Mudit Soni , Surbhi Agrawal This is my paper · ORCID
classification cs.CVcs.AI
keywords long-tailedrecognitionlogitadjustmenteffectivepriorclassimbalancepost-hoccorrectionBayesianposteriordistributionalignment
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

The paper sets out to show that the class bias of a long-tailed classifier is better captured by the model's own effective prior — the average of its predicted probabilities over the data — than by the empirical class frequencies used in standard logit adjustment. It proposes Prior2Posterior (P2P), a post-hoc reweighting of the predicted posterior by the ratio of the balanced test prior to this learned prior, and proves that the adjusted posterior has exactly the balanced test marginals for both plain cross-entropy models and models trained with logit-adjusted loss. The authors argue that this makes the correction optimal for the balanced test set, and their experiments show consistent accuracy gains over class-frequency adjustment on CIFAR-LT, ImageNet-LT, and iNaturalist18. The practical claim is that P2P can be dropped onto already-trained models of many existing methods and improve them without retraining.

What carries the argument

The load-bearing object is the effective prior $P^m(y)$, defined as the integral of the model's softmax posterior over the data distribution and estimated in practice by averaging $P^m(y|x)$ over training samples (Eq. 27) and, for logit-adjusted models, over a balanced validation set or a frequency-scaled training estimate (Eqs. 28–29). This prior carries the entire correction: each predicted posterior is multiplied by $P^t(y)/P^m(y)$, which down-weights classes the model over-predicts and up-weights classes it under-predicts, so the aggregate predictions are forced to match the balanced test prior. The same estimated prior is what lets the method inspect any trained model for residual bias and remove it post-hoc.

What would settle it

On a two-class Gaussian toy problem with imbalance 100, train a linear model, compute the Bayes-optimal decision boundary, and then apply the P2P correction; if the corrected boundary and balanced accuracy differ from the Bayes classifier even though the adjusted class proportions are balanced, then marginal matching is not the same as optimality.

Watch

Extended reading notes

Core claim

The central discovery is that the prior implicit in a trained model, $P^m(y)=\int P^m(y|x)P(x)\,dx$, differs systematically from the class-frequency prior, so frequency-based corrections remove the wrong amount of bias. The paper derives the adjusted posterior $P^a(y|x)=P^m(y|x)\frac{P^t(y)}{P^m(y)}\frac{P(x)}{P^t(x)}$ for plain cross-entropy training (Theorem 1) and $P^a(y|x)=P^m(y|x)\frac{P^t(y)}{P^m(y)}$ for logit-adjusted training (Theorem 2), where $P^m(y)$ is the effective prior estimated from the model's own outputs. In both cases the adjustment forces the aggregate predicted class proportions on the test distribution to equal the balanced prior $P^t(y)$, which the paper identifies as the defining optimality condition. Empirically this corrected posterior improves accuracy on long-tailed benchmarks and, applied post-hoc, boosts many existing methods beyond their original results.

Load-bearing premise

The argument collapses if matching the predicted class marginal to the balanced test prior is not sufficient for accuracy optimality, since the correction never changes the per-sample ordering of classes that actually decides each prediction.

Editorial extensions

If this is right

  • P2P applied to a plain cross-entropy model beats class-frequency logit adjustment on CIFAR10-LT, CIFAR100-LT, ImageNet-LT, and iNaturalist18.
  • Applied to logit-adjusted models (classifier retraining or feature tuning), P2P removes residual head-class bias and gives further gains, with the best results from feature tuning plus P2P.
  • Many existing published methods improve when P2P is applied to their pretrained outputs, with reported gains up to 1.48 points on ImageNet-LT and 1.67 points on iNaturalist18 without retraining.
  • P2P also tracks test-time prior shifts better than distribution-alignment baselines across forward and backward imbalance ratios.
  • Averaging the training-set and validation-set estimates of the effective prior yields the best correction, indicating that the estimate improves with more samples.

Reading between the lines

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

  • Beyond the paper: the proof establishes marginal matching, $P^a(y)=P^t(y)$, not per-sample Bayes optimality; two posteriors with identical marginals can have different error rates, so the "optimality" claim is stronger than what the theorem shows.
  • Beyond the paper: because the correction targets $P^t(y)$, the same recipe can point at any desired test prior, so it could be extended to deliberately imbalanced or shifted test distributions beyond the ones tested.
  • Beyond the paper: the consistent gains on pretrained models suggest P2P may partly act as a calibration step; comparing it against temperature scaling or per-class calibration on the same models would separate prior correction from general calibration effects.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes a post-hoc logit-adjustment method for long-tailed recognition, called Prior2Posterior (P2P). Rather than using class frequencies to model the training class prior, the authors estimate an 'effective model prior' P^m(y) from the model's own average posterior predictions over the training (or validation) data and then adjust predictions by multiplying by P^t(y)/P^m(y), with an additional P(x)/P^t(x) factor in the plain-cross-entropy case. The paper claims to prove optimality of this adjustment for both plain cross-entropy training (Theorem 1) and logit-adjusted training (Theorem 2). Experiments cover toy data, CIFAR10/100-LT, ImageNet-LT, and iNaturalist-18, showing consistent gains over several baselines and small improvements when the correction is applied post-hoc to existing methods.

Significance. The empirical observation that the model's own marginal predictions can differ from class frequencies, and using that to correct residual bias, is a plausible and practically useful idea. The experimental breadth is a strength: the method is tested on several large-scale benchmarks, test-time shifted distributions, and applied to a variety of existing methods, with gains that are consistently positive. However, the central theoretical claim is not what it appears to be. The paper's 'optimality' is defined as matching the test class marginal, not as accuracy or Bayes-risk optimality, and the marginal-matching property is proven only for an unnormalized density that does not correspond to the softmax-normalized classifier actually deployed. The empirical results, while suggestive, are mostly small and do not by themselves rescue the broken theory.

major comments (3)
  1. [Section 4, Definition (Eq. 7) and Theorem 1 (Eqs. 8-13)] The optimality claim is definitional. The Definition states that P^a(y|x) is 'optimal' if its test class marginal satisfies P^a(y)=P^t(y); Theorem 1 then proves exactly this marginal property for the unnormalized quantity in Eq. (8). No accuracy or decision-theoretic optimality is established. The phrasing 'optimal adjustment' in the Theorem statement and Section 4.1 therefore overstates what is proven; at most, this is a marginal-matching adjustment.
  2. [Section 4.1, Eq. (8) and proof Eqs. (10)-(13)] The marginal-matching property does not survive softmax normalization, which is the normalization actually used in the deployed classifier. The adjusted unnormalized score is P^a(y|x)=P^m(y|x) P^t(y)/P^m(y) * P(x)/P^t(x), and the deployed rule is Q(y|x)=P^a(y|x)/\sum_z P^a(z|x). The class-independent factor P(x)/P^t(x) cancels in this ratio, so Q(y|x)=P^m(y|x) P^t(y)/P^m(y) / \sum_z P^m(z|x) P^t(z)/P^m(z), and the integral \int Q(y|x) P^t(x) dx is not in general equal to P^t(y). The proof in Eqs. (10)-(13) integrates the unnormalized P^a and never normalizes over classes. For a concrete failure, take two classes A,B, P^t(x) uniform on two points, P(x=1)=0.9, P^m(A|x=1)=0.9, P^m(A|x=2)=0.1, and P^t(A)=P^t(B)=0.5; then the normalized marginal is Q(A)≈0.344, not 0.5. Thus Theorem 1's guarantee does not hold for the prediction rule used in the experiments.
  3. [Section 4.3, Eq. (27)-(29)] The 'effective model prior' is estimated from the model's own predicted posterior, and the correction then divides by that same estimated quantity. This makes the correction essentially a self-referential rescaling: the 'prior' is derived from the posterior it is meant to correct, and the theorem's conclusion is enforced by construction on the unnormalized scores. The claim that the estimated P^m(y) captures a true underlying bias requires the additional assumption that the model's conditional probabilities are calibrated estimates of P(y|x), an assumption that is neither established nor plausibly satisfied given the paper's own discussion of overfitting to head classes. The empirical gains should therefore be interpreted as evidence for a heuristic, not as confirmation of the theoretical optimality claim.
minor comments (4)
  1. [Section 4.2, before Theorem 2] The notation for the two distributions in the logit-adjusted case is inconsistent: 'P m(y|x)' is used both for the inference-time model and for the training-time adjusted distribution. Please introduce distinct symbols (e.g., P^m and \tilde P^m) and use them consistently throughout the proof of Theorem 2.
  2. [Table 3] The reported improvements are all positive, but no error bars or significance tests are given. Given the small margins on ImageNet-LT (e.g., +0.22 to +0.35 for several methods), it would be useful to state whether these differences exceed run-to-run variation.
  3. [Section 6.4] The text says 'As shown in Table 8' but the table with the full test-time-shift results appears in the supplementary material; please provide a cross-reference that distinguishes the main-text table from the supplementary table.
  4. [Eq. (28)] The normalization in Eq. (28) uses the same denominator \sum_k n_k as Eq. (27), but the average is over validation samples from P^t(x); the correct normalization constant should be the number of validation samples, not the training-set size. This should be stated explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1's 'optimality' claim is the Definition itself: Eq. (8) is constructed to make Eq. (7) hold, and the softmax-normalized rule used in experiments is not covered by the proof.

  1. self definitional [Section 4, Definition (Eq. 7); Theorem 1 (Eq. 8) and Proof (Eqs. 10-13); Theorem 2 (Eqs. 15-26)]
    "The adjusted distribution P a(y|x) is optimal for the test dataset with marginal distributions P t(x) and P t(y) if it satisfies following property P a(y) = ∫ P a(y|x)P t(x)dx = P t(y) (7) ... Theorem 1. ... P a(y|x) = P m(y|x) P t(y)/P m(y) P (x)/P t(x) (8) ... Proof. ... = P t(y) (13)"

    Optimality is defined as the very property that Theorem 1 then proves. Substituting Eq. (8) into the integral in Eq. (10) cancels P(x) and P^t(x) and returns P^t(y) by pure algebra; the formula was chosen so this cancellation occurs. The theorem therefore establishes no independent optimality criterion such as minimum classification error or Bayes-optimal decisions; it merely restates the defining marginal-matching property. Theorem 2 repeats the same construction for the logit-adjusted model. Moreover, the paper calls P(x)/P^t(x) inconsequential because it is absorbed by normalization, but under softmax normalization that class-independent factor cancels, so the proved marginal property does not automatically hold for the posterior actually used to make predictions.

full rationale

The primary theoretical result is circular in the sense defined by the review rubric. Section 4 defines an adjustment to be 'optimal' exactly when its test marginal equals P^t(y) (Eq. 7), then Theorem 1 presents a formula (Eq. 8) and proves, by substituting that formula into the defining integral, that the marginal equals P^t(y) (Eqs. 10-13). This is a construction rather than a derivation of optimality with respect to accuracy or Bayes risk. The same structure is repeated in Theorem 2 for logit-adjusted training. The empirical results on CIFAR-LT, ImageNet-LT, and iNaturalist18 provide genuine external support for the correction as a heuristic, and the only self-citation ([2]) is a related-work mention, not load-bearing. Because the central 'optimality' proof is definitional rather than independent, the circularity score is 6.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The central correction relies on one fitted scalar alpha, the class-conditional distribution invariance assumption, the assumption that softmax outputs are usable posterior estimates, and the paper-specific definition that optimality equals marginal matching. No new physical entities are introduced; the 'effective prior' is a derived statistic from the model's own outputs.

free parameters (1)
  • alpha (logit adjustment scalar)
    Tuned on holdout validation per dataset to scale the estimated prior, as in Menon et al.; values not reported in the paper.
assumptions (3)
  • domain assumption Class-conditional generative distributions are identical between train and test: P(x|y)=P^t(x|y).
    Stated in Section 3.2 and used to derive Eq. 3 and both theorems.
  • domain assumption DNN softmax outputs approximate Bayesian posterior probabilities P(y|x).
    Invoked in Sections 1 and 3.2 via citations [14,30]; load-bearing because P^m(y) is defined as the marginal of these outputs.
  • ad hoc to paper A post-hoc adjustment is optimal iff its class marginal on the test distribution equals P^t(y).
    Definition in Section 4. This equates optimality with marginal matching, a much weaker condition than accuracy optimality.
invented entities (1)
  • Effective model prior P^m(y)
    purpose: Quantifies the class marginal the trained model actually induces, used as the base for post-hoc logit reweighting.
    Defined as the average of the model's own softmax outputs over training/validation data. It is a model-derived statistic with no independent falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prior2Posterior: Model Prior Correction for Long-Tailed Learning." pith.science (2026). https://pith.science/paper/BCQXY6KH

@misc{pith2026241216540,
  author       = {Pith},
  title        = {Pith review of: Prior2Posterior: Model Prior Correction for Long-Tailed Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCQXY6KH}},
  note         = {Machine review of arXiv:2412.16540}
}
read the original abstract

Learning-based solutions for long-tailed recognition face difficulties in generalizing on balanced test datasets. Due to imbalanced data prior, the learned \textit{a posteriori} distribution is biased toward the most frequent (head) classes, leading to an inferior performance on the least frequent (tail) classes. In general, the performance can be improved by removing such a bias by eliminating the effect of imbalanced prior modeled using the number of class samples (frequencies). We first observe that the \textit{effective prior} on the classes, learned by the model at the end of the training, can differ from the empirical prior obtained using class frequencies. Thus, we propose a novel approach to accurately model the effective prior of a trained model using \textit{a posteriori} probabilities. We propose to correct the imbalanced prior by adjusting the predicted \textit{a posteriori} probabilities (Prior2Posterior: P2P) using the calculated prior in a post-hoc manner after the training, and show that it can result in improved model performance. We present theoretical analysis showing the optimality of our approach for models trained with naive cross-entropy loss as well as logit adjusted loss. Our experiments show that the proposed approach achieves new state-of-the-art (SOTA) on several benchmark datasets from the long-tail literature in the category of logit adjustment methods. Further, the proposed approach can be used to inspect any existing method to capture the \textit{effective prior} and remove any residual bias to improve its performance, post-hoc, without model retraining. We also show that by using the proposed post-hoc approach, the performance of many existing methods can be improved further.

Figures

Figures reproduced from arXiv: 2412.16540 by the authors.

Figure 1
Figure 1. We present the results on a toy dataset with the imbalance factor of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We show marginal class probabilities, P(y), for classes on CIFAR100-LT dataset with imbalance factor of 200. The ef￾fective prior calculated using proposed approach and using class frequencies are shown for head classes (Many-first column), tail classes (Few-third column) and rest of the classes (Medium￾middle column). One may note that model shows bias towards the head classes. Further, the bias is under represente… view at source ↗
Figure 3
Figure 3. The performance on Imagenet-LT (a) for Stage 1 (CE) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The performance on iNaturalist18 for Stage 1 baseline [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Proposed approach is summarized in the figure. The top row illustrates plain CE loss and logit-adjusted loss showing that model [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 51 canonical work pages

  1. [1]

    Long-tailed recognition via weight balancing

    Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In IEEE CVPR, pages 6897–6907, 2022. 3, 5, 7, 2

    work page 2022

  2. [2]

    Robust loss function for class imbalanced semantic seg- mentation and image classification

    S Divakar Bhat, Amit More, Mudit Soni, and Yuji Yasui. Robust loss function for class imbalanced semantic seg- mentation and image classification. IFAC-PapersOnLine, 56(2):7934–7939, 2023. 3

    work page 2023

  3. [3]

    A systematic study of the class imbalance problem in convo- lutional neural networks

    Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convo- lutional neural networks. Neural networks , 106:249–259,

    work page

  4. [4]

    Learning imbalanced datasets with label- distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label- distribution-aware margin loss. Advances in NeurIPS , 32,

    work page

  5. [5]

    Area: Adaptive reweight- ing via effective area for long-tailed classification

    Xiaohua Chen, Yucan Zhou, Dayan Wu, Chule Yang, Bo Li, Qinghua Hu, and Weiping Wang. Area: Adaptive reweight- ing via effective area for long-tailed classification. In IEEE ICCV, pages 19277–19287, 2023. 5, 7

    work page 2023

  6. [6]

    Reslt: Residual learning for long-tailed recogni- tion

    Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Reslt: Residual learning for long-tailed recogni- tion. IEEE TPAMI, 45(3):3695–3706, 2022. 5, 6, 7, 2

    work page 2022

  7. [7]

    Class-balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In IEEE CVPR, pages 9268–9277, 2019. 3, 5, 7

    work page 2019

  8. [8]

    Global and local mixture consistency cumulative learning for long-tailed visual recognitions

    Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recognitions. In IEEE CVPR, pages 15814–15823, 2023. 3, 6

    work page 2023

Show all 56 references
  1. [9]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In Proceedings of ICML, pages 1321–1330. PMLR, 2017. 4

    2017

  2. [10]

    Disentangling label dis- tribution for long-tailed visual recognition

    Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label dis- tribution for long-tailed visual recognition. In IEEE CVPR, pages 6626–6636, 2021. 3, 5, 7, 2

    2021

  3. [11]

    Learning deep representation for imbalanced classi- fication

    Chen Huang, Yining Li, Chen Change Loy, and Xiaoou Tang. Learning deep representation for imbalanced classi- fication. In IEEE CVPR, pages 5375–5384, 2016. 3

    2016

  4. [12]

    Decou- pling representation and classifier for long-tailed recogni- tion

    Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decou- pling representation and classifier for long-tailed recogni- tion. arXiv preprint arXiv:1910.09217 , 2019. 2, 3, 5, 6, 7

    1910 arXiv

  5. [13]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7

    2009

  6. [14]

    Neural network classification and prior class probabilities

    Steve Lawrence, Ian Burns, Andrew Back, Ah Chung Tsoi, and C Lee Giles. Neural network classification and prior class probabilities. In Neural networks: tricks of the trade , pages 299–313. Springer, 2002. 1, 2, 4

    2002

  7. [15]

    Long-tailed visual recognition via gaussian clouded logit adjustment

    Mengke Li, Yiu-ming Cheung, and Yang Lu. Long-tailed visual recognition via gaussian clouded logit adjustment. In IEEE CVPR, pages 6929–6938, 2022. 2

    2022

  8. [16]

    Metasaug: Meta seman- tic augmentation for long-tailed visual recognition

    Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta seman- tic augmentation for long-tailed visual recognition. In IEEE CVPR, pages 5212–5221, 2021. 6

    2021

  9. [17]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. In IEEE ICCV, pages 2980–2988, 2017. 3

    2017

  10. [18]

    Inducing neural collapse in deep long- tailed learning

    Xuantong Liu, Jianfeng Zhang, Tianyang Hu, He Cao, Yuan Yao, and Lujia Pan. Inducing neural collapse in deep long- tailed learning. In AISTATS, pages 11534–11544. PMLR,

  11. [19]

    Large-scale long-tailed recognition in an open world

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In IEEE CVPR, pages 2537– 2546, 2019. 7

    2019

  12. [20]

    Curvature-balanced feature manifold learning for long-tailed classification

    Yanbiao Ma, Licheng Jiao, Fang Liu, Shuyuan Yang, Xu Liu, and Lingling Li. Curvature-balanced feature manifold learning for long-tailed classification. In IEEE CVPR, pages 15824–15835, 2023. 3, 5, 7

    2023

  13. [21]

    knn approach to unbalanced data distributions: a case study involving information extrac- tion

    Inderjeet Mani and I Zhang. knn approach to unbalanced data distributions: a case study involving information extrac- tion. In Proceedings of workshop on learning from imbal- anced datasets, volume 126, pages 1–7. ICML, 2003. 2

    2003

  14. [22]

    When does imbalanced data require more than cost-sensitive learning

    Dragos Margineantu. When does imbalanced data require more than cost-sensitive learning. In Proceedings of the AAAI Workshop on Learning from Imbalanced Data Sets , pages 47–50, 2000. 1, 2

    2000

  15. [23]

    Long-tail learning via logit adjustment

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 2, 3, 6

    2007 arXiv

  16. [24]

    Decoupled training for long-tailed classification with stochastic representations

    Giung Nam, Sunguk Jang, and Juho Lee. Decoupled training for long-tailed classification with stochastic representations. arXiv preprint arXiv:2304.09426, 2023. 5, 7, 2

    2023 arXiv

  17. [25]

    The majority can help the minority: Context-rich minority oversampling for long-tailed classifi- cation

    Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classifi- cation. In IEEE CVPR, pages 6887–6896, 2022. 6

    2022

  18. [26]

    Feature directions matter: Long-tailed learning via rotated balanced represen- tation

    Gao Peifeng, Qianqian Xu, Peisong Wen, Zhiyong Yang, Huiyang Shao, and Qingming Huang. Feature directions matter: Long-tailed learning via rotated balanced represen- tation. 2023. 5, 7, 2

    2023

  19. [27]

    Optimal transport for long-tailed recognition with learnable cost matrix

    Hanyu Peng, Mingming Sun, and Ping Li. Optimal transport for long-tailed recognition with learnable cost matrix. In In- ternational conference on learning representations, 2021. 2

    2021

  20. [28]

    Escaping saddle points for effective generalization on class- imbalanced data

    Harsh Rangwani, Sumukh K Aithal, Mayank Mishra, et al. Escaping saddle points for effective generalization on class- imbalanced data. Advances in NeurIPS , 35:22791–22805,

  21. [29]

    Balanced meta-softmax for long-tailed visual recogni- tion

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recogni- tion. Advances in NeurIPS, 33:4175–4186, 2020. 2, 3, 5, 7, 1

    2020

  22. [30]

    Neural net- work classifiers estimate bayesian a posteriori probabilities

    Michael D Richard and Richard P Lippmann. Neural net- work classifiers estimate bayesian a posteriori probabilities. Neural computation, 3(4):461–483, 1991. 2, 4

    1991

  23. [31]

    Distributional robustness loss for long-tail learning

    Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In IEEE ICCV , pages 9495–9504,

  24. [32]

    Relative entropic optimal transport: a (prior-aware) match- ing perspective to (unbalanced) classification

    Liangliang Shi, Haoyu Zhen, Gu Zhang, and Junchi Yan. Relative entropic optimal transport: a (prior-aware) match- ing perspective to (unbalanced) classification. Advances in Neural Information Processing Systems, 36, 2024. 2

    2024

  25. [33]

    Meta-weight-net: Learn- ing an explicit mapping for sample weighting

    Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learn- ing an explicit mapping for sample weighting. Advances in NeurIPS, 32, 2019. 5

    2019

  26. [34]

    Rethinking the inception ar- chitecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception ar- chitecture for computer vision. In IEEE CVPR, pages 2818– 2826, 2016. 4

    2016

  27. [35]

    Long- tailed classification by keeping the good and removing the bad momentum causal effect

    Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long- tailed classification by keeping the good and removing the bad momentum causal effect. Advances in NeurIPS , 33:1513–1524, 2020. 7, 3

    2020

  28. [36]

    Improving tail-class representation with centroid contrastive learning

    Anthony Meng Huat Tiong, Junnan Li, Guosheng Lin, Boyang Li, Caiming Xiong, and Steven CH Hoi. Improving tail-class representation with centroid contrastive learning. Pattern Recognition Letters, 168:123–130, 2023. 2

    2023

  29. [37]

    The inaturalist species classification and de- tection dataset

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and de- tection dataset. In IEEE CVPR, pages 8769–8778, 2018. 7

    2018

  30. [38]

    Solar: Sinkhorn la- bel refinery for imbalanced partial-label learning

    Haobo Wang, Mingxuan Xia, Yixuan Li, Yuren Mao, Lei Feng, Gang Chen, and Junbo Zhao. Solar: Sinkhorn la- bel refinery for imbalanced partial-label learning. Advances in neural information processing systems , 35:8104–8117,

  31. [39]

    Rsg: A simple but effective mod- ule for learning imbalanced datasets

    Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective mod- ule for learning imbalanced datasets. In IEEE CVPR, pages 3784–3793, 2021. 2

    2021

  32. [40]

    Seesaw loss for long- tailed instance segmentation

    Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long- tailed instance segmentation. In IEEE CVPR, pages 9695– 9704, 2021. 3

    2021

  33. [41]

    Adaptive class suppression loss for long-tail object detection

    Tong Wang, Yousong Zhu, Chaoyang Zhao, Wei Zeng, Jin- qiao Wang, and Ming Tang. Adaptive class suppression loss for long-tail object detection. In IEEE CVPR, pages 3103– 3112, 2021. 3

    2021

  34. [42]

    Long-tailed recognition by rout- ing diverse distribution-aware experts

    Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by rout- ing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020. 6, 8

    2010 arXiv

  35. [43]

    Margin calibration for long-tailed visual recognition

    Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jin- dong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition. In ACML, pages 1101–1116. PMLR, 2023. 3, 5, 7, 2

    2023

  36. [44]

    Learn- ing to model the tail

    Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learn- ing to model the tail. Advances in NeurIPS, 30, 2017. 3

    2017

  37. [45]

    Distribution alignment: A unified framework for long-tail visual recognition

    Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In IEEE CVPR , pages 2361– 2370, 2021. 3, 5, 7, 2

    2021

  38. [46]

    Im- proving calibration for long-tailed recognition

    Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Im- proving calibration for long-tailed recognition. In IEEE CVPR, pages 16489–16498, 2021. 6, 8

    2021

  39. [47]

    Class-conditional sharpness-aware mini- mization for deep long-tailed recognition

    Zhipeng Zhou, Lanqing Li, Peilin Zhao, Pheng-Ann Heng, and Wei Gong. Class-conditional sharpness-aware mini- mization for deep long-tailed recognition. In IEEE CVPR, pages 3499–3509, 2023. 3, 5, 7, 2

    2023

  40. [48]

    Balanced contrastive learning for long-tailed visual recognition

    Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In IEEE CVPR, pages 6908– 6917, 2022. 6, 8 Prior2Posterior: Model Prior Correction for Long-Tailed Learning Supplementary Material

    2022

  41. [49]

    In Stage 2, for both classifier and feature tuning cases, we train the models for 4000 iterations

    Additional training details We train Stage 1 models using cross-entropy loss for CIFAR10-LT and CIFAR100-LT datasets for 20,000 itera- tions. In Stage 2, for both classifier and feature tuning cases, we train the models for 4000 iterations. For ImageNet-LT dataset, we train th...

  42. [50]

    For the sake of complete- ness we provide a mathematical justification using the ap- proach of moment matching

    Dissimilarity of data distributions: P (X) ̸= P t(X) In the paper we clearly maintain the distributions P (X) and P t(X) as distinct in nature. For the sake of complete- ness we provide a mathematical justification using the ap- proach of moment matching. In particular, we sho...

  43. [51]

    From the figure it is clearly observed that model bias is quite different from empirical bias estimated using class frequen- cies

    The effective prior for ImageNet-LT and iNaturalist18 datasets In Figure 4 we show the model bias estimated using class frequencies and the effective prior calculated using pro- posed approach on ImageNet-LT and iNaturalist18 datasets. From the figure it is clearly observed th...

  44. [52]

    We note from the table that proposed approach achieves highest overall accuracy while shot-wise accuracies are not affected much

    Multishot accuracies In Table 6 and Table 7 we show multi-shot accuracies for ImageNet-LT and iNaturalist18 datasets and compare it with some of the recently published methods. We note from the table that proposed approach achieves highest overall accuracy while shot-wise accu...

  45. [53]

    8 we compare model performance for test-time shifted distributions with additional baselines and a few more distribution shifts

    Additional results on test time shifted im- balance In Table. 8 we compare model performance for test-time shifted distributions with additional baselines and a few more distribution shifts. We note the superior performance of proposed algorithm

  46. [54]

    The performance on iNaturalist18 for Stage 1 baseline (CE) and Stage 2 (CL and FT) are shown

    Discussion on Distribution Matching Recent works like [27] have proposed to tackle this dis- tribution misalignment problem from an optimisation per- Figure 5. The performance on iNaturalist18 for Stage 1 baseline (CE) and Stage 2 (CL and FT) are shown. The effect of post-hoc ...

  47. [55]

    The manually specified smooth- ing guidance matrix Q can be seen as a generic representa- tion for the effective prior

    present a novel variant of the optimal transport called, Relative Entropic Optimal Transport to learn match- ing with a specified prior. The manually specified smooth- ing guidance matrix Q can be seen as a generic representa- tion for the effective prior

  48. [56]

    The block diagram illustrates the different stages involved in the process starting from bias accumulation in traditional training to bias removal us- ing the proposed method

    Flow of the proposed approach We summarise the proposed approach in a block dia- gram as shown in Figure6. The block diagram illustrates the different stages involved in the process starting from bias accumulation in traditional training to bias removal us- ing the proposed me...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.