Pith. sign in

REVIEW 4 major objections 5 minor 74 references

BAPE: Learning an Explicit Bayes Classifier for Long-tailed Visual Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read BAPE learns an explicit Bayes classifier for long-tailed visual recognition by estimating class-conditional feature distributions directly, replacing gradient descent on the classifier with point estimation and improving accuracy over…

desk verdict A simple, mostly sound point-estimation variant of the vMF classifier with modest but consistent gains; the Bayes-optimality claim outruns the evidence until the distributional assumption is checked. read the letter →

arxiv 2506.23280 v1 pith:IRHF5H7J submitted 2025-06-29 cs.LG

classification cs.LG MSC 62F1562H1168T07
keywords long-tailedrecognitionBayesclassifiervonMises-FisherdistributionpointestimationadjustmentlogitclassimbalanceMAP
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 tries to establish that a Bayes-optimal classifier for long-tailed visual recognition can be learned explicitly, without gradient descent on the classifier, by modeling feature distributions and estimating their parameters from data. It argues that existing methods implicitly estimate posterior probabilities through losses like softmax cross-entropy, which suffer from gradient imbalance and fail to ensure the Bayes decision rule under class imbalance. BAPE assumes features of each class follow a von Mises-Fisher distribution on the unit sphere, estimates the mean direction and concentration via maximum a posteriori estimation, and then plugs these estimates into Bayes' theorem to form a linear classifier. The paper also introduces a distribution-adjustment step that rescales the concentration parameters to adapt the classifier to a test set with arbitrary class imbalance. If correct, this would offer a simpler, training-stable alternative to logit adjustment and related post-hoc methods, with consistent gains on CIFAR-10/100-LT, ImageNet-LT, and iNaturalist 2018.

What carries the argument

The central object is the von Mises-Fisher (vMF) distribution on the unit sphere, used to model class-conditional feature densities f_p(z|μ_y, κ_y) = exp(κ_y μ_y^T z)/C_p(κ_y). The argument is carried by the MAP estimator of the vMF parameters: the posterior mean direction is the (regularized) sample mean and the concentration parameter satisfies I_{p/2}(κ)/I_{p/2-1}(κ) = β/α, with a simple closed-form approximation κ ≈ pβ/($α^{2}$ - $β^{2}$). These estimates, combined with an equiangular tight-frame prior on the mean directions, yield an explicit Bayes classifier that is a linear function of the feature z. The distribution-adjustment step rescales κ_y according to the target class prior, which is what lets the classifier transfer from the training imbalance to the test imbalance. The machinery thus replaces gradient-based classifier learning with moment-based estimation, which is what the paper claims alleviates gradient imbalance and guarantees the Bayes decision rule.

What would settle it

Train a backbone on a long-tailed dataset, then record the empirical class-conditional feature distribution after the BAPE projection head. Compute the angular variance per class and compare it to the variance predicted by the MAP-estimated κ_y under the vMF model; a systematic mismatch (e.g., features concentrated in several clusters rather than one) would show the distributional assumption fails. Alternatively, create a test set where the class frequencies are deliberately different from both the training prior and the adjusted prior used by BAPE, and check whether accuracy on the tail classes tracks the adjusted prior; a failure to adapt would contradict the claimed distribution-adjustment benefit.

Watch

Extended reading notes

Core claim

The central claim is that the posterior probability p(y|z) can be explicitly constructed as p(y) times a von Mises-Fisher density over the feature embedding z, with class-specific mean direction μ_y and concentration κ_y, and that these parameters can be estimated by maximum a posteriori point estimation rather than learned by gradient descent. Given features normalized to the unit sphere and a prior over the classifier parameters, the posterior takes a closed form whose mean direction is the sample mean and whose concentration satisfies a ratio of Bessel functions. The paper shows this MAP estimator can be computed online from first sample moments, and that the resulting classifier—a linear classifier in feature space—avoids the gradient-imbalance problem that plagues implicit softmax training. A distribution-adjustment technique then rescales κ_y to match a target imbalance factor for the test set, allowing the classifier trained on a long-tailed training set to be adapted to a balanced or differently imbalanced test distribution without retraining. Empirically, the method improves top-1 accuracy over Logit Adjustment by 1.6 to 2.0 points on CIFAR-100-LT, 1.6 points on ImageNet-LT, and 1.3 points on iNaturalist 2018.

Load-bearing premise

The load-bearing premise is that, after the projection head and normalization, the features of each class genuinely follow a von Mises-Fisher distribution with a single concentration parameter κ_y; if real features have a different shape (multimodal, non-isotropic, or with varying concentration), the estimated 'Bayes optimal' classifier is not actually optimal and reduces to a fixed linear classifier whose parameters are estimated from sample moments.

Editorial extensions

If this is right

  • If the central claim holds, long-tailed classifiers can be trained with a backbone plus a simple moment-estimated head, eliminating the need for careful classifier-side gradient balancing and reducing sensitivity to minority-class gradient suppression.
  • The distribution-adjustment technique implies that a single trained model can be re-targeted to any test-time class frequency by changing only the κ_y parameters, enabling deployment without retraining when the test distribution shifts in imbalance.
  • Because the BAPE head is linear and explicit, it can be combined with any representation-learning or backbone-training method, so its gains are orthogonal to existing long-tailed losses and can be stacked on top of them.
  • The explicit distributional form gives a principled way to diagnose when a feature space is poorly calibrated for Bayes-optimal decisions: if the vMF assumption is violated, the estimated κ will not reflect the true angular concentration, and the classifier's decisions will deviate from the Bayes rule.

Reading between the lines

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

  • Editorial extension: The paper's reliance on the vMF assumption suggests a testable diagnostic—if real class-conditional features are multimodal or have varying concentration across the sphere, BAPE's fixed linear classifier should underperform a learned classifier on the same backbone; this is a concrete condition under which the claimed Bayes optimality would fail.
  • Editorial extension: The distribution-adjustment idea could be extended to other exponential-family feature distributions (e.g., Gaussian with class-specific covariance), where the post-hoc rescaling would be a covariance or precision adjustment; the paper does not explore this, but the same point-estimation logic would apply.
  • Editorial extension: The paper reports gains mainly on top of Logit Adjustment; a natural follow-up is to measure BAPE's benefit on top of representation-learning methods like supervised contrastive learning, where the vMF shape may be more or less accurate than under softmax-trained backbones.
  • Editorial extension: The claim that gradient imbalance is alleviated because no gradient flows through the classifier is plausible, but the backbone still trains with gradients; a direct test would be to measure the effective per-class gradient norms and minority-class feature collapse during BAPE training versus standard softmax training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes BAPE, a long-tailed visual recognition method that models L2-normalized feature embeddings of each class as a von Mises-Fisher (vMF) distribution, estimates the class mean directions and concentration parameters by maximum a posteriori point estimation with an ETF-based prior, and plugs these estimates into the Bayes classifier formula (Eq. 6). The BAPE classifier is trained jointly with a logit-adjustment (LA) classifier for stable early training, and a 'distribution adjustment' is applied to adapt the classifier to an arbitrary test imbalance factor. Experiments on CIFAR-10/100-LT, ImageNet-LT, and iNaturalist 2018 report consistent top-1 accuracy gains over a Logit Adjustment baseline.

Significance. If the vMF assumption and the claimed Bayes-optimality guarantee held, BAPE would be a conceptually interesting departure from gradient-trained classifiers, with a transparent point-estimation procedure and a cheap post-hoc adaptation mechanism. The empirical gains over Logit Adjustment are consistent across four benchmarks, and the ablation study gives some support for the prior and distribution-adjustment components. However, the paper does not currently substantiate the central distributional assumption, leaves the distribution-adjustment rule unspecified, and overstates the 'without gradient descent' contribution. These issues concern the paper's main theoretical claims rather than the empirical comparison, so the manuscript needs substantial revision before the claims can be accepted as stated.

major comments (4)
  1. [§3.2, Eq. (6)] The Bayes-optimality claim is conditional on the vMF assumption, but the assumption is not validated. Eq. (6) is exactly the posterior of a generative vMF model; if the actual projected, L2-normalized features are anisotropic, multimodal, or exhibit within-class heterogeneity, the sample-moment estimates in §3.3 do not yield the Bayes classifier, and BAPE reduces to a fixed linear classifier estimated by first moments. The supporting evidence in §3.2 is the SELI-based motivation and the t-SNE/ellipse visualization in Fig. 2, neither of which is a goodness-of-fit test. Please add distributional diagnostics (for example, per-class comparison of the empirical mean resultant length with the vMF prediction, or a simulation-based calibration check) or explicitly reframe the theoretical claim as 'Bayes optimal under the vMF model'.
  2. [§3.3–§4.4, Eq. (10), Tabs. 5–6] The distribution adjustment is never defined mathematically. The text states that the method adjusts parameters to match a test set with an arbitrary imbalance factor and that κ is the key quantity, but no equation specifies how κ_y or μ_y are modified as a function of the target class frequencies. Tabs. 5 and 6 show that this step has a substantial effect (e.g., 52.5 versus 50.9 with and without DA), so the missing definition is not a presentation issue; it is required for reproducibility. Please state the adjustment rule explicitly, including whether it acts on κ, on μ, or on the prior probabilities, and whether it is applied only at test time or also during training.
  3. [Abstract, §3.3, §3.4, Eq. (13), Conclusion] The claim that BAPE 'learns the Bayes classifier without gradient descent' is overstated. Only the classifier parameters κ_y and μ_y are obtained by point estimation in closed form; the backbone and projection head are trained by backpropagation through LBAPE, Eq. (13) additionally trains an LA classifier by gradient descent, and §3.3 states that the prior directions m0 are themselves updated by gradient descent during early training. This matters because the abstract and conclusion tie the method's contribution to eliminating gradient imbalance. Please replace the wording with a precise statement, e.g., 'the BAPE classifier parameters are obtained by point estimation, while the feature extractor and the auxiliary branch are still trained by gradient descent'.
  4. [§3.4, §4.3] The orthogonality claim is not supported by the reported experiments. The paper says that LLA can be replaced with any off-the-shelf long-tailed method and that BAPE's gains are orthogonal to existing approaches, but all main results use LA as the auxiliary branch (Tabs. 1–3). Adding at least one result with a different auxiliary loss (e.g., LDAM, class-balanced loss, or a re-sampling baseline) would substantiate the claim; as written, the empirical contribution is specifically 'BAPE on top of LA'.
minor comments (5)
  1. [Eq. (10)] The κ approximation is written as pβα/(α^2−β^2), which is easy to misread as pβα; please typeset it as \frac{p\beta\alpha}{\alpha^2-\beta^2} and note that it is the large-κ limit of κ ≈ pR/(1−R^2), where R = β/α.
  2. [Tab. 5] The meaning of the check marks in the 'Training' and 'Testing' columns is unclear, especially for the LA rows; please add a legend or a more descriptive caption.
  3. [§4.2] The prior hyperparameters α̂0 and β̂0 are set to different values for each dataset without a sensitivity analysis or a stated selection criterion; please report a small sensitivity study or explain how these values were chosen.
  4. [Throughout] There are minor language and consistency issues: 'BAPE surpass Logit Adj.' should be 'BAPE surpasses Logit Adj.', and the dataset name appears as both 'CIFAR-10/100-LT' and 'CIFAR-LT-10/100'.
  5. [Fig. 1] The figure caption says the plot shows 'κ (BAPE) and the norm ... in LA', but the axes are not labeled and the visual claim that BAPE is 'frequency independent' is not established by the plot; please label the axes and clarify what is plotted for each class.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BAPE's Bayes classifier is a derivation from an explicit vMF assumption with parameters estimated from training data, and the reported gains are external benchmark evaluations.

full rationale

The derivation chain in Sec. 3.2-3.3 is a model-based calculation rather than a circular reduction. Eq. (6) states the posterior p(y|z) that follows from Bayes' theorem once each class conditional is assumed vMF with parameters (mu_y, kappa_y); this is an explicit distributional assumption, not an output of the method that is then relabeled as a prediction. The parameters are estimated from training features via MAP (Eqs. 7-11) and no test labels are used. The central empirical claims are comparisons against Logit Adjustment and other published methods on CIFAR-LT, ImageNet-LT, and iNaturalist, so the reported improvements are not forced by construction. The self-citations in the introduction are contextual and are not used to justify the vMF assumption, the MAP estimator, or the optimality claim. Concerns that the vMF shape may not hold for real features, and that per-dataset hyperparameters (alpha0, beta0) are tuned, are correctness and robustness issues rather than circularity; they do not make Eq. (6) equivalent to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the vMF distribution assumption and on the per-dataset prior hyperparameters; no invented entities are introduced. The method also depends on the transferability of the conditional feature distribution from training to test for the distribution adjustment.

free parameters (2)
  • prior strength alpha0 = 40 (CIFAR-LT), 20 (ImageNet-LT), 10 (iNaturalist 2018)
    Chosen per dataset to control the influence of the vMF prior; no stated selection rule, so it acts as a tunable hyperparameter.
  • prior length beta0 = 8 (CIFAR-LT), 0.6 (ImageNet-LT), 0.3 (iNaturalist 2018)
    Set per dataset with alpha0; affects the MAP estimate of kappa and hence classifier behavior.
assumptions (4)
  • domain assumption Class-conditional features follow a vMF distribution on the unit sphere
    Introduced in Sec 3.2 and used in Eq (6) to derive the Bayes classifier form; not empirically validated in the paper beyond feature visualizations.
  • standard math The conjugate prior for vMF parameters has the stated form (Eq 7)
    Standard result in directional statistics; used to derive MAP estimates in Sec 3.3.
  • standard math The MAP equations (Eq 9) and the approximation kappa_hat (Eq 10) are valid
    Standard results from [53]; the displayed formula in the submitted text appears garbled, so the exact approximation used is ambiguous.
  • domain assumption The conditional feature distribution transfers from training to test, so only the class prior or kappa need adjustment
    Underlies the distribution adjustment technique described in Sec 3.2 and analyzed in Tab. 5; if the shift is more complex, the adjustment is unjustified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BAPE: Learning an Explicit Bayes Classifier for Long-tailed Visual Recognition." pith.science (2026). https://pith.science/paper/IRHF5H7J

@misc{pith2026250623280,
  author       = {Pith},
  title        = {Pith review of: BAPE: Learning an Explicit Bayes Classifier for Long-tailed Visual Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRHF5H7J}},
  note         = {Machine review of arXiv:2506.23280}
}
read the original abstract

Bayesian decision theory advocates the Bayes classifier as the optimal approach for minimizing the risk in machine learning problems. Current deep learning algorithms usually solve for the optimal classifier by \emph{implicitly} estimating the posterior probabilities, \emph{e.g.}, by minimizing the Softmax cross-entropy loss. This simple methodology has been proven effective for meticulously balanced academic benchmark datasets. However, it is not applicable to the long-tailed data distributions in the real world, where it leads to the gradient imbalance issue and fails to ensure the Bayes optimal decision rule. To address these challenges, this paper presents a novel approach (BAPE) that provides a more precise theoretical estimation of the data distributions by \emph{explicitly} modeling the parameters of the posterior probabilities and solving them with point estimation. Consequently, our method directly learns the Bayes classifier without gradient descent based on Bayes' theorem, simultaneously alleviating the gradient imbalance and ensuring the Bayes optimal decision rule. Furthermore, we propose a straightforward yet effective \emph{distribution adjustment} technique. This method enables the Bayes classifier trained from the long-tailed training set to effectively adapt to the test data distribution with an arbitrary imbalance factor, thereby enhancing performance without incurring additional computational costs. In addition, we demonstrate the gains of our method are orthogonal to existing learning approaches for long-tailed scenarios, as they are mostly designed under the principle of \emph{implicitly} estimating the posterior probabilities. Extensive empirical evaluations on CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist demonstrate that our method significantly improves the generalization performance of popular deep networks, despite its simplicity.

Figures

Figures reproduced from arXiv: 2506.23280 by the authors.

Figure 2
Figure 2. Visualization of the feature embedding via t-SNE. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 62 canonical work pages

  1. [1]

    A systematic study of the class imbalance problem in convolutional neural networks

    Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018

  2. [2]

    What is the effect of importance weighting in deep learning? In ICML, 2019

    Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In ICML, 2019

  3. [3]

    Domain balancing: Face recognition on long-tailed domains

    Dong Cao, Xiangyu Zhu, Xingyu Huang, Jianzhu Guo, and Zhen Lei. Domain balancing: Face recognition on long-tailed domains. In CVPR, 2020

  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. In NeurIPS, 2019

  5. [5]

    SMOTE: synthetic minority over-sampling technique

    Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: synthetic minority over-sampling technique. JAIR, 2002

  6. [6]

    AutoAugment: Learning augmentation strategies from data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. AutoAugment: Learning augmentation strategies from data. In CVPR, 2019

  7. [7]

    RandAugment: Practical automated data augmentation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. RandAugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, 2020

  8. [8]

    ResLT: Residual learning for long-tailed recognition

    Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. ResLT: Residual learning for long-tailed recognition. TPAMI, 2022

Show all 74 references
  1. [9]

    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 CVPR, 2019

  2. [10]

    The mnist database of handwritten digit images for machine learning research

    Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 2012

  3. [11]

    Improved regularization of convolutional neural networks with cutout

    Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint, 2017

  4. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

  5. [13]

    Switta: Switching domain experts and aggregating contextual features towards realistic test-time adaptation

    Chaoqun Du, Jiayi Guo, Yulin Wang, and Gao Huang. Switta: Switching domain experts and aggregating contextual features towards realistic test-time adaptation. In ICML Workshop, 2025

  6. [14]

    Simpro: A simple probabilistic framework towards realistic long-tailed semi-supervised learning

    Chaoqun Du, Yizeng Han, and Gao Huang. Simpro: A simple probabilistic framework towards realistic long-tailed semi-supervised learning. arXiv preprint arXiv:2402.13505, 2024

  7. [15]

    Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation

    Chaoqun Du, Yulin Wang, Jiayi Guo, Yizeng Han, Jie Zhou, and Gao Huang. Unitta: Unified benchmark and versatile framework towards realistic test-time adaptation. arXiv preprint arXiv:2407.20080, 2024

  8. [16]

    Probabilistic contrastive learning for long-tailed visual recognition

    Chaoqun Du, Yulin Wang, Shiji Song, and Gao Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 10

  9. [17]

    Exploring deep neural networks via layer- peeled model: Minority collapse in imbalanced training

    Cong Fang, Hangfeng He, Qi Long, and Weijie J Su. Exploring deep neural networks via layer- peeled model: Minority collapse in imbalanced training. the National Academy of Sciences, 2021

  10. [18]

    Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016

  11. [19]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In NeurIPS, 2014

  12. [20]

    Assessing a single image in reference-guided image synthesis

    Jiayi Guo, Chaoqun Du, Jiangshan Wang, Huijuan Huang, Pengfei Wan, and Gao Huang. Assessing a single image in reference-guided image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 753–761, 2022

  13. [21]

    Faceclip: Facial image-to-video translation via a brief text description

    Jiayi Guo, Hayk Manukyan, Chenyu Yang, Chaofei Wang, Levon Khachatryan, Shant Navasardyan, Shiji Song, Humphrey Shi, and Gao Huang. Faceclip: Facial image-to-video translation via a brief text description. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4...

  14. [22]

    Zero-shot generative model adaptation via image-specific prompt learning

    Jiayi Guo, Chaofei Wang, You Wu, Eric Zhang, Kai Wang, Xingqian Xu, Shiji Song, Humphrey Shi, and Gao Huang. Zero-shot generative model adaptation via image-specific prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11...

  15. [23]

    Smooth diffusion: Crafting smooth latent spaces in diffusion models

    Jiayi Guo, Xingqian Xu, Yifan Pu, Zanlin Ni, Chaofei Wang, Manushree Vasu, Shiji Song, Gao Huang, and Humphrey Shi. Smooth diffusion: Crafting smooth latent spaces in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages...

  16. [24]

    Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic-domain alignment

    Jiayi Guo, Junhao Zhao, Chaoqun Du, Yulin Wang, Chunjiang Ge, Zanlin Ni, Shiji Song, Humphrey Shi, and Gao Huang. Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic-domain alignment. In Proceedings of the Computer Vision and Pattern Recognition Co...

  17. [25]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016

  18. [26]

    Disentangling label distribution for long-tailed visual recognition

    Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In CVPR, 2021

  19. [27]

    Learning deep representation for imbalanced classification

    Chen Huang, Yining Li, Chen Change Loy, and Xiaoou Tang. Learning deep representation for imbalanced classification. In CVPR, 2016

  20. [28]

    The high separation probability assumption for semi-supervised learning

    Gao Huang and Chaoqun Du. The high separation probability assumption for semi-supervised learning. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 52(12):7561–7573, 2022

  21. [29]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017

  22. [30]

    Decoupling representation and classifier for long-tailed recognition

    Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020

  23. [31]

    Striking the right balance with uncertainty

    Salman Khan, Munawar Hayat, Syed Waqas Zamir, Jianbing Shen, and Ling Shao. Striking the right balance with uncertainty. In CVPR, 2019

  24. [32]

    Label-imbalanced and group-sensitive classification under overparameterization

    Ganesh Ramachandra Kini, Orestis Paraskevas, Samet Oymak, and Christos Thrampoulidis. Label-imbalanced and group-sensitive classification under overparameterization. NeurIPS, 2021

  25. [33]

    Learning multiple layers of features from tiny images.(2009), 2009

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images.(2009), 2009. 11

  26. [34]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, 2017

  27. [35]

    Addressing the curse of imbalanced training sets: one-sided selection

    Miroslav Kubat, Stan Matwin, et al. Addressing the curse of imbalanced training sets: one-sided selection. In ICML, 1997

  28. [36]

    Nested collaborative learning for long-tailed visual recognition

    Jun Li, Zichang Tan, Jun Wan, Zhen Lei, and Guodong Guo. Nested collaborative learning for long-tailed visual recognition. In CVPR, 2022

  29. [37]

    MetaSAug: Meta semantic augmentation for long-tailed visual recognition

    Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. MetaSAug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, 2021

  30. [38]

    Self supervision to distillation for long-tailed visual recognition

    Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In ICCV, 2021

  31. [39]

    Targeted supervised contrastive learning for long-tailed recognition

    Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In CVPR, 2022

  32. [40]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In ICCV, 2017

  33. [41]

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Open long-tailed recognition in a dynamic world. TPAMI, 2022

  34. [42]

    Exploring the limits of weakly supervised pretraining

    Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018

  35. [43]

    Directional statistics

    Kanti V Mardia, Peter E Jupp, and KV Mardia. Directional statistics. Wiley Online Library, 2000

  36. [44]

    Long-tailed visual recognition via gaussian clouded logit adjustment

    Yang Lu Mengke Li, Yiu-ming Cheung. Long-tailed visual recognition via gaussian clouded logit adjustment. In CVPR, 2022

  37. [45]

    On the statistical consistency of algorithms for binary classification under class imbalance

    Aditya Menon, Harikrishna Narasimhan, Shivani Agarwal, and Sanjay Chawla. On the statistical consistency of algorithms for binary classification under class imbalance. In ICML, 2013

  38. [46]

    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. In ICLR, 2021

  39. [47]

    Machine learning: a probabilistic perspective

    Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012

  40. [48]

    Machine learning from imbalanced data sets 101

    Foster Provost. Machine learning from imbalanced data sets 101. In AAAI’2000 workshop, 2000

  41. [49]

    Balanced meta-softmax for long-tailed visual recognition

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, and Shuai Yi. Balanced meta-softmax for long-tailed visual recognition. In NeurIPS, 2020

  42. [50]

    Imagenet large scale visual recognition challenge

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015

  43. [51]

    Distributional robustness loss for long-tail learning

    Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In ICCV, 2021

  44. [52]

    Deep unsuper- vised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynamics. In ICML, 2015

  45. [53]

    A short note on parameter approximation for von mises-fisher distributions: and a fast implementation of is(x)

    Suvrit Sra. A short note on parameter approximation for von mises-fisher distributions: and a fast implementation of is(x). Computational Statistics, 2012

  46. [54]

    Equalization loss for long-tailed object recognition

    Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 12

  47. [55]

    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. In NeurIPS, 2020

  48. [56]

    Imbalance trouble: Revisiting neural-collapse geometry

    Christos Thrampoulidis, Ganesh Ramachandra Kini, Vala Vakilian, and Tina Behnia. Imbalance trouble: Revisiting neural-collapse geometry. In NeurIPS, 2022

  49. [57]

    A simple framework for contrastive learning of visual representations

    Chen Ting, Kornblith Simon, Norouzi Mohammad, and Hinton Geoffrey. A simple framework for contrastive learning of visual representations. In ICML, 2020

  50. [58]

    Visualizing data using t-SNE

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. JMLR, 2008

  51. [59]

    The inaturalist species classification and detection 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 detection dataset. In CVPR, 2018

  52. [60]

    Principles of risk minimization for learning theory

    Vladimir Vapnik. Principles of risk minimization for learning theory. NeurIPS, 1991

  53. [61]

    The nature of statistical learning theory

    Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999

  54. [62]

    Class imbalance, redux

    Byron C Wallace, Kevin Small, Carla E Brodley, and Thomas A Trikalinos. Class imbalance, redux. In ICDM, 2011

  55. [63]

    CosFace: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large margin cosine loss for deep face recognition. In CVPR, 2018

  56. [64]

    Towards calibrated hyper-sphere representation via distribution overlap coefficient for long- tailed learning

    Hualiang Wang, Siming Fu, Xiaoxuan He, Hangxiang Fang, Zuozhu Liu, and Haoji Hu. Towards calibrated hyper-sphere representation via distribution overlap coefficient for long- tailed learning. In ECCV, 2022

  57. [65]

    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 CVPR, 2021

  58. [66]

    Long-tailed recognition by routing diverse distribution-aware experts

    Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2020

  59. [67]

    Distribution-balanced loss for multi-label classification in long-tailed datasets

    Tong Wu, Qingqiu Huang, Ziwei Liu, Yu Wang, and Dahua Lin. Distribution-balanced loss for multi-label classification in long-tailed datasets. In ECCV, 2020

  60. [68]

    Aggregated residual transformations for deep neural networks

    Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017

  61. [69]

    Do we really need a learnable classifier at the end of deep neural network? NeurIPS, 2022

    Yibo Yang, Liang Xie, Shixiang Chen, Xiangtai Li, Zhouchen Lin, and Dacheng Tao. Do we really need a learnable classifier at the end of deep neural network? NeurIPS, 2022

  62. [70]

    Rethinking the value of labels for improving class-imbalanced learning

    Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In NeurIPS, 2020

  63. [71]

    Identifying and compensat- ing for feature deviation in imbalanced deep learning

    Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensat- ing for feature deviation in imbalanced deep learning. arXiv preprint, 2020

  64. [72]

    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 CVPR, 2021

  65. [73]

    Adaptive logit adjustment loss for long-tailed visual recognition

    Yan Zhao, Weicong Chen, Xu Tan, Kai Huang, and Jihong Zhu. Adaptive logit adjustment loss for long-tailed visual recognition. In AAAI, 2022

  66. [74]

    BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition

    Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, 2020. 13

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.