Pith. sign in

REVIEW 3 major objections 5 minor 39 references

A power-law logit adjustment, Class-Balanced Softmax, generalizes balanced softmax and improves test accuracy on long-tailed benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:18 UTC pith:CBDJ2B2H

load-bearing objection A simple beta-tuned extension of Balanced Softmax with a useful diagnostic and broad experiments, but the significance claims on two main benchmarks don't match the reported means. the 3 major comments →

arxiv 2607.22258 v1 pith:CBDJ2B2H submitted 2026-07-24 cs.LG cs.CV

Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition

Yi-Hang Zhu , Rajeev Raman , Shiqi Su , Jianyuan Sun , Xinyu Yang , Nan Xing , Huiyu Zhou This is my paper
classification cs.LG cs.CV
keywords long-tailed recognitionimbalanced classificationlogit adjustmentsoftmaxBayesian decision theorypreference issuemodel imbalance levelpower-law approximation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a single power-law exponent on class frequencies is a principled correction for class-imbalanced training, and that the standard balanced-softmax fix is a rigid special case of this correction. The central derivation starts from a Bayesian class-probability shift theorem: if one model uses the standard softmax and a second uses a class-weighted softmax, the second is equivalent to the first with logits shifted by the log-ratio of the two class-probability distributions. Assuming these distributions are power laws in class size gives the CBS logit adjustment, adding β log |N_c| to each logit, with β≥0. The paper also names and quantifies a 'preference issue'—higher training error and larger generalization gaps for tail classes—via a model imbalance metric I, and reports experiments showing CBS lowers I and improves all-class test recall relative to Balanced Softmax and many other rebalancing methods on ImageNet-LT, Places-LT, iNaturalist2018, and a 1203-class LVIS-based dataset. A sympathetic reader would care because CBS is simple, computationally free at inference, and offers a single knob to trade off head and tail performance.

Core claim

At its core, the paper claims that rebalancing methods should not assume the class-probability distribution of a well-trained model equals the empirical class frequency. Theorem 4.1 shows that if a model is trained with a softmax reweighted by \p(y=c)/p(y=c), its predictions are equivalent to the original model with logits shifted by log(\p(y=c)/p(y=c)). The paper then approximates both p(y=c) for a balanced model and \p(y=c) for an imbalanced model as power laws in |N_c|, with exponents of opposite sign, yielding the CBS softmax \phi_nc ∝ |N_c|^β e^{z_nc}. This is the paper's central object: a one-parameter family of logit adjustments that contains Balanced Softmax (β=1) and, for β≠1, can s

What carries the argument

The load-bearing object is the class-probability shift theorem (Theorem 4.1) combined with two power-law approximations: p(y=c) ∝ |N_c|^β_b with β_b≤0 and \p(y=c) ∝ |N_c|^β_ib with β_ib≥0. The theorem expresses the conditional probability of a model trained with a modified softmax as a Gibbs distribution with logits shifted by log(\p(y=c)/p(y=c)). Substituting the power-law forms yields the CBS logit adjustment z_nc + β log |N_c|, β≥0, which the paper shows rebalances gradient rewards and penalties across head and tail classes. The model imbalance level I, defined as the range of group-wise preference scores, serves as the diagnostic that tracks the preference issue CBS is designed to mitiga

Load-bearing premise

The derivation depends on the heuristic assumption that both the balanced model's and the imbalanced model's class-probability distributions are power-law functions of the training class count; the paper itself acknowledges this is not a proven universal property, and if the true shift is not a power law, the Bayesian argument does not single out CBS as the optimal correction.

What would settle it

Measure, for a long-tailed dataset, the actual class-probability distributions p(y=c) of a balanced model and \p(y=c) of an imbalanced model using target probabilities on training data; compute the ratio r_c = \p(y=c)/p(y=c). If a logit adjustment using log r_c outperforms CBS with the best-tuned β by a practically meaningful margin—or if the power-law fit to r_c is obviously poor while CBS still wins—that would show whether the power-law approximation is the decisive factor or merely one convenient functional form.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • CBS is a drop-in replacement for the softmax layer in any classifier; it adds no trainable parameters and only a class-count-dependent logit shift, so it is trivially integrated into existing pipelines.
  • Because β can be tuned (the paper finds values in [1.1, 1.3] work well on the large benchmarks), CBS gives practitioners a single scalar to control the head-tail trade-off, in contrast to the fixed β=1 of Balanced Softmax.
  • CBS composes with contrastive representation learning: combining CBS with BCL or ProCo yields higher all-class test recall than those methods with Balanced Softmax, indicating the rebalancing effect is complementary to representation improvements.
  • On extremely imbalanced and high-class-count datasets (iNaturalist2018 with 8,142 classes, the LVIS-based task with 1,203 classes and a 50,550:1 imbalance), CBS's gains over Balanced Softmax remain statistically significant, suggesting the advantage does not vanish at scale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the power-law approximation is a good model, then the multiplicative form |N_c|^β suggests that the 'right' rebalancing strength may depend on dataset-specific factors such as intra-class variance and data quality; estimating β from data (e.g., by maximizing held-out likelihood) could replace the current grid search without changing the architecture.
  • The preference-issue framing implies that rebalancing is needed even when the test distribution is unknown or imbalanced; this could make CBS more robust than methods that explicitly assume a balanced test set, which is relevant for open-world recognition and detection.
  • The same Bayesian shift theorem can generate other logit adjustments by choosing different parametric families for p(y=c) and \p(y=c); comparing power laws against, say, saturating functions of class size would test whether the functional form or the logit-shift principle drives the observed gains.
  • A direct falsification route: on a dataset where the empirical ratio \p(y=c)/p(y=c) deviates strongly from a power law in |N_c|, CBS should underperform a logit shift computed directly from the measured ratio; an experiment along these lines would cleanly separate the theorem's validity from the power-law heuristic.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Class-Balanced Softmax (CBS), a logit-adjustment training loss for long-tailed recognition. CBS multiplies each class logit by |N_c|^β before the softmax, generalising Balanced Softmax (which is the β=1 special case). The method is motivated by a Bayesian change-of-measure argument (Theorem 4.1) combined with a power-law approximation of class-prior shifts (Eqs. 24–25), which yields the closed-form adjustment in Eq. (27). The paper also introduces a 'preference issue' — higher training error and larger generalisation gap for tail classes — quantified by a model imbalance level I. Experiments are reported on long-tailed CIFAR-10/100, ImageNet-LT, Place-LT, iNaturalist2018, and an LVIS-based dataset, with statistical tests comparing CBS against Balanced Softmax.

Significance. If fully supported, CBS would be a simple, computationally cheap improvement over a widely used baseline, and the proposed model imbalance level could be a useful diagnostic for long-tailed training. The paper is transparent about its limitations, states its assumptions clearly, and provides a code link. However, the theoretical contribution is weakened by the heuristic power-law assumption and by the fact that β is a free parameter selected on test accuracy. More importantly, the empirical evidence as presented contains internal statistical inconsistencies that, once corrected, do not support the claim of significant gains on at least one of the main large-scale benchmarks. The central empirical claim therefore needs careful revision.

major comments (3)
  1. [Section 5.2 and Table 4] The statistical reporting is internally inconsistent. For ImageNet-LT, Table 6 gives Balanced Softmax 51.06±0.82 and CBS 51.56±0.65, a mean difference of 0.50, not Δμ=1.27 as stated in Figure 4. Recomputing Welch's t-test from these summary statistics gives t≈1.07, df≈7.6, p≈0.32, and Cohen's d≈0.68, not p=0.0278 and d=1.51. For Place-LT, Table 7 gives 29.42±0.45 vs 30.11±0.45, a difference of 0.69, not 1.23; this yields p≈0.04 and d≈1.53, not p=0.0005 and d=3.24. In addition, the C100-LT results in Appendix E are admitted non-significant (p≈0.46, 0.19, 0.10). Therefore the statement in §5.3.2 that 'CBS achieves significantly higher (p-value<0.05) average testing recall across all classes compared to Balanced Softmax' is not supported by the reported summary statistics. Please correct the test statistics, or supply per-run data that justify them, and temper the significance claims to the
  2. [Section 5.2 and Table 4]
  3. [Eqs. (24)–(25) and Conclusion]
minor comments (5)
  1. [Section 5.2] Typo: 'SDG' should be 'SGD' in the description of the optimiser.
  2. [Eq. (26)] The approximation sign is unnecessary: the step from the previous expression to Eq. (26) is exact because the normalisation sums over c' are constants that cancel between numerator and denominator. This is a presentation issue, but the current '≈' is misleading.
  3. [Section 3 after Theorem 3.1] The statement that 'as training progresses, the class-imbalanced Δgrad converges to zero, while the model increasingly exhibits the preference issue' is not proved. Please mark it as a conjecture or provide supporting evidence.
  4. [Figure 6] The colour scale and the 'Clipped Gradient Balance Intensity' label are difficult to interpret. Please clarify how the values are clipped and what the colour intensity represents.
  5. [Table 4] The darker-intensity highlighting is not explained in the caption. Also, since β values for other datasets are not reported, consider adding an appendix table with the selected β for every dataset.

Circularity Check

0 steps flagged

No significant circularity: the Bayesian-to-logit derivation is algebraic, the power-law ansatz is disclosed as heuristic, and the reported benchmark gains are empirical claims rather than definitional reductions.

full rationale

The derivation chain is self-contained and does not reduce to its own inputs. Theorem 4.1 is a standard Bayes-form softmax shift (built on the external Balanced Softmax paper [12]), and Eqs. (21)-(27) are arithmetic given the stated power-law ansatz. The paper explicitly labels the power-law assumption as heuristic: in Section 6 it says the power-law functions 'are only supported empirically in the studied settings. They are not established as a universal property of all long-tailed recognition problems.' This is an admitted modeling assumption, not a hidden circularity. The hyperparameter β is not derived from the theory; Section 5.2 says 'The preferred value of β for each dataset is determined via a grid search.' Consequently, the claim that CBS outperforms Balanced Softmax rests on external benchmark experiments, not on a derivation that assumes that conclusion. The preference-issue metric I is a new measurement, and CBS is not defined in terms of I, so there is no self-definitional loop. No load-bearing self-citations, uniqueness-theorem import, or ansatz smuggled via prior work are present. The statistical inconsistencies between Figure 4 and Tables 6-7 (e.g., ImageNet-LT Δμ reported as 1.27 vs. 0.50 from the table means; Place-LT 1.23 vs. 0.69) are correctness risks, not circularity, and do not change the circularity verdict.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

The central derivation adds two power-law approximations and a tuned β that are not derived from first principles; Theorem 4.1 is a standard reweighted-softmax identity. The new conceptual entity (preference issue) is measurable, but the CBS formula itself follows from heuristic assumptions plus a fitted exponent.

free parameters (2)
  • β (class-frequency exponent in CBS) = dataset-specific; 1.3 chosen for Place-LT; full per-dataset values not reported
    Controls the strength of the logit adjustment in Eq. (27) and is selected by grid search on test accuracy (Table 4), so the central method depends on a fitted, dataset-dependent constant.
  • β_b and β_ib (power-law exponents for class priors) = not separately identified; only β = β_ib − β_b is used
    Introduced in Eqs. (24)-(25) to approximate p(y=c) and \hat p(y=c); the decomposition is non-identifiable and functions as a theoretical placeholder rather than an estimated quantity.
axioms (3)
  • ad hoc to paper p(y=c) and \hat p(y=c) can be approximated as powers of class frequency: κ_c=|N_c|^{β_b}, \hatκ_c=|N_c|^{β_ib} (Eqs. 24-25).
    Adopted heuristically in §4; the paper's limitations state the power-law form is not established as universal, so this is an assumption introduced to make the CBS formula tractable.
  • domain assumption For an ideally balanced model, the target-class posterior is 1−ϵ and all wrong classes share ϵ/(|C|−1) (Appendix C).
    Used to prove Theorem 4.2; supported empirically by Figure F.8 but not proven to hold for all balanced long-tailed models.
  • standard math The exponential-family parametrisation of softmax and the canonical link relation used in Theorem 4.1 are standard.
    The proof in Appendix B is built on the standard softmax link, following [12].
invented entities (1)
  • Preference issue independent evidence
    purpose: Named phenomenon: imbalanced models show higher training error and larger generalization gap for tail classes; used to motivate CBS and to define the model imbalance level I.
    Operationalized through Eqs. (10)-(12) and observed in Figure 1 independently of CBS, so it has falsifiable, measurable content.

pith-pipeline@v1.3.0-alltime-deepseek · 20307 in / 21099 out tokens · 191532 ms · 2026-08-01T05:18:54.106322+00:00 · methodology

0 comments
read the original abstract

Deep learning models using traditional softmax classifiers have achieved remarkable success in various classification tasks. However, their performance degrades significantly on imbalanced datasets. Although Balanced Softmax is widely adopted as a state-of-the-art rebalancing method, it possesses inherent limitations, such as yielding disproportionately lower testing accuracy for tail classes. To mitigate these shortcomings, we propose the Class-Balanced Softmax (CBS). Rooted in a theoretical Bayesian framework and a heuristic power-law assumption, the CBS is a simple logit adjustment that is computationally inexpensive and easily integrated into existing pipelines. Furthermore, we characterise a fundamental phenomenon in models trained on imbalanced data, termed the preference issue, wherein models exhibit higher training error and a larger generalisation gap for classes with limited data. To quantify this issue, we introduce a novel metric and demonstrate that CBS effectively mitigates the preference issue. Extensive experiments on large-scale benchmarks show that CBS is highly scalable and outperforms existing methods, including Balanced Softmax.

Figures

Figures reproduced from arXiv: 2607.22258 by Huiyu Zhou, Jianyuan Sun, Nan Xing, Rajeev Raman, Shiqi Su, Xinyu Yang, Yi-Hang Zhu.

Figure 1
Figure 1. Figure 1: The preference issue observed in the model trained on ImageNet-LT: significantly higher training [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Class probability distribution generated by a balanced model on the ImageNet-LT training data. [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantitative comparison of the model imbalance level [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Statistical comparison of the average testing recall across all classes for Balanced Softmax and [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Statistical comparison of the average testing recall across all classes for Balanced Softmax and [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The higher values indicate greater emphasis by the model regarding gradients during training. [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references

  1. [1]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Ieee, 2009, pp. 248–255

    2009

  2. [2]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep con- volutional neural networks, in: Proceedings of Neural Information Processing Systems (NeurIPS), V ol. 25, 2012

    2012

  3. [3]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  4. [4]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations (ICLR), 2021

    2021

  5. [5]

    K. He, G. Gkioxari, P. Dollár, R. Girshick, Mask r-cnn, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2961– 2969

    2017

  6. [6]

    C.-Y . Wang, A. Bochkovskiy, H.-Y . M. Liao, Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 7464–7475

    2023

  7. [7]

    Van Horn, O

    G. Van Horn, O. Mac Aodha, Y . Song, Y . Cui, C. Sun, A. Shepard, H. Adam, P. Perona, S. Belongie, The inaturalist species classification and detection dataset, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8769–8778

    2018

  8. [8]

    Gupta, P

    A. Gupta, P. Dollar, R. Girshick, Lvis: A dataset for large vocabulary instance segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5356–5364. 35

    2019

  9. [9]

    Zhang, B

    Y . Zhang, B. Kang, B. Hooi, S. Yan, J. Feng, Deep long-tailed learning: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)

    2023

  10. [10]

    L. Yang, H. Jiang, Q. Song, J. Guo, A survey on long-tailed visual recognition, International Journal of Computer Vision 130 (7) (2022) 1837–1872

    2022

  11. [11]

    K. Cao, C. Wei, A. Gaidon, N. Arechiga, T. Ma, Learning imbalanced datasets with label-distribution-aware margin loss, in: Proceedings of Neural Information Processing Systems (NeurIPS), 2019

    2019

  12. [12]

    J. Ren, C. Yu, S. Sheng, X. Ma, H. Zhao, S. Yi, H. Li, Balanced meta-softmax for long-tailed visual recognition, in: Proceedings of Neural Information Processing Systems (NeurIPS), 2020

    2020

  13. [13]

    Zhang, C

    E. Zhang, C. Geng, S. Chen, Class-aware universum inspired re-balance learning for long-tailed recognition, Pattern Recognition 161 (2025) 111337

    2025

  14. [14]

    X. Wang, L. Lian, Z. Miao, Z. Liu, S. Yu, Long-tailed recognition by routing di- verse distribution-aware experts, in: International Conference on Learning Rep- resentations (ICLR), 2021

    2021

  15. [15]

    J. Cui, Z. Zhong, Z. Tian, S. Liu, B. Yu, J. Jia, Generalized parametric con- trastive learning, IEEE Transactions on Pattern Analysis and Machine Intelli- gence (2023)

    2023

  16. [16]

    Z. Tan, J. Li, J. Du, J. Wan, Z. Lei, G. Guo, Ncl++: Nested collaborative learning for long-tailed visual recognition, Pattern Recognition 147 (2024) 110064

    2024

  17. [17]

    W. Liu, Z. Wu, Y . Wang, H. Ding, F. Liu, J. Lin, G. Lin, Lcreg: Long-tailed image classification with latent categories based recognition, Pattern Recognition 145 (2024) 109971

    2024

  18. [18]

    N. V . Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, Smote: synthetic minority over-sampling technique, Journal of Artificial Intelligence Research 16 (2002) 321–357. 36

    2002

  19. [19]

    Zhang, M

    H. Zhang, M. Cisse, Y . N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, in: International Conference on Learning Representations (ICLR), 2018

    2018

  20. [20]

    Xiang, J

    L. Xiang, J. Han, G. Ding, Margin-aware rectified augmentation for long-tailed recognition, Pattern Recognition 141 (2023) 109608

    2023

  21. [21]

    J. S. Baik, I. Y . Yoon, J. W. Choi, Dbn-mix: Training dual branch network using bilateral mixup augmentation for long-tailed visual recognition, Pattern Recogni- tion 147 (2024) 110107

    2024

  22. [22]

    Y . Yang, S. Chen, X. Li, L. Xie, Z. Lin, D. Tao, Inducing neural collapse in imbal- anced learning: Do we really need a learnable classifier at the end of deep neural network?, in: Proceedings of Neural Information Processing Systems (NeurIPS), V ol. 35, 2022, pp. 37991–38002

    2022

  23. [23]

    Peifeng, Q

    G. Peifeng, Q. Xu, P. Wen, Z. Yang, H. Shao, Q. Huang, Feature directions mat- ter: Long-tailed learning via rotated balanced representation, in: International Conference on Machine Learning (ICML), PMLR, 2023

    2023

  24. [24]

    B. Kang, S. Xie, M. Rohrbach, Z. Yan, A. Gordo, J. Feng, Y . Kalantidis, Decou- pling representation and classifier for long-tailed recognition, in: International Conference on Learning Representations (ICLR), 2020

    2020

  25. [25]

    Z. Liu, Z. Miao, X. Zhan, J. Wang, B. Gong, S. X. Yu, Large-scale long-tailed recognition in an open world, in: IEEE Conference on Computer Vision and Pat- tern Recognition (CVPR), 2019

    2019

  26. [26]

    Zhou, X.-Y

    Z.-H. Zhou, X.-Y . Liu, Training cost-sensitive neural networks with methods ad- dressing the class imbalance problem, IEEE Transactions on Knowledge and Data Engineering 18 (1) (2005) 63–77

    2005

  27. [27]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980–2988. 37

    2017

  28. [28]

    A. K. Menon, S. Jayasumana, A. S. Rawat, H. Jain, A. Veit, S. Kumar, Long-tail learning via logit adjustment, in: International Conference on Learning Repre- sentations (ICLR), 2021

    2021

  29. [29]

    J. Cui, S. Liu, Z. Tian, Z. Zhong, J. Jia, Reslt: Residual learning for long-tailed recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (3) (2023) 3695–3706

    2023

  30. [30]

    Bengio, A

    Y . Bengio, A. Courville, P. Vincent, Representation learning: A review and new perspectives, IEEE Transactions on Pattern Analysis and Machine Intelligence 35 (8) (2013) 1798–1828

    2013

  31. [31]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, A. Courville, Deep learning, MIT press, 2016

    2016

  32. [32]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton, et al., Learning multiple layers of features from tiny images (2009)

    2009

  33. [33]

    A. C. Wilson, R. Roelofs, M. Stern, N. Srebro, B. Recht, The marginal value of adaptive gradient methods in machine learning, in: Proceedings of Neural Infor- mation Processing Systems (NeurIPS), V ol. 30, 2017

    2017

  34. [34]

    J. Zhu, Z. Wang, J. Chen, Y .-P. P. Chen, Y .-G. Jiang, Balanced contrastive learning for long-tailed visual recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 6908–6917

    2022

  35. [35]

    C. Du, Y . Wang, S. Song, G. Huang, Probabilistic contrastive learning for long- tailed visual recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  36. [36]

    Loshchilov, F

    I. Loshchilov, F. Hutter, SGDR: Stochastic gradient descent with warm restarts, in: International Conference on Learning Representations (ICLR), 2017

    2017

  37. [37]

    H. Yan, Y . Qian, F. Peng, J. Luo, F. Li, et al., Neural collapse to multiple centers for imbalanced data, in: Proceedings of Neural Information Processing Systems (NeurIPS), 2024. 38

    2024

  38. [38]

    J. Gao, H. Zhao, D. dan Guo, H. Zha, Distribution alignment optimization through neural collapse for long-tailed classification, in: International Conference on Ma- chine Learning (ICML), 2024

    2024

  39. [39]

    T.-Y . Pan, C. Zhang, Y . Li, H. Hu, D. Xuan, S. Changpinyo, B. Gong, W.-L. Chao, On model calibration for long-tailed object detection and instance seg- mentation, Proceedings of Neural Information Processing Systems (NeurIPS) 34 (2021) 2529–2542. 39

    2021