Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Long-Tailed Learning for Generalized Category Discovery

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims long-tailed category discovery is best handled by a two-stage framework: learnable-distribution pseudo-labels plus neighborhood-based representation balancing, reporting state-of-the-art accuracy on four benchmarks.

desk verdict A credible two-stage method for long-tailed GCD, but the representation balancing stage needs a collapse check before the performance gains can be trusted. read the letter →

arxiv 2506.06965 v1 pith:CRJJEB5T submitted 2025-06-08 cs.AI cs.CV

Cuong Manh Hoang This is my paper
classification cs.AIcs.CV
keywords generalizedcategorydiscoverylong-tailedlearningpseudo-labelingSinkhorn-Knopprepresentationbalancingneighborhooddensityclassimbalanceclustering
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 argues that existing generalized category discovery (GCD) methods fail on long-tailed data — imbalanced data where a few head classes dominate and many tail classes have few samples — for two identifiable reasons: pseudo-labeling with a uniform target distribution pushes head-class samples into tail classes, and unsupervised contrastive learning stays biased toward dense head regions. To fix the first, the framework estimates the class distribution through long-tailed clustering and uses it to guide a learnable distribution that feeds the Sinkhorn-Knopp pseudo-labeling step over a large queue. To fix the second, a representation balancing stage measures each sample's neighborhood density, treats low-density samples as tail classes, and trains the backbone to pull samples toward their local neighborhood means with a density-weighted loss. On four long-tailed benchmarks the method reports overall test accuracies of 95.0%, 74.4%, 88.4%, and 32.5%, surpassing the previous state-of-the-art by 3.9, 7.2, 4.7, and 2.6 percentage points. If those numbers hold, long-tailed GCD is not an inherently harder problem but one whose machinery must be distribution-aware in both labeling and representation learning.

What carries the argument

The load-bearing object is the balanced loss in Eq. (10), $L_{\text{bal}} = \frac{1}{B}\sum_i (1+w_i)(1-\mathrm{sim}(z_i,\mu_i))$, where $\mu_i$ is the mean of $z_i$ and its $K$ nearest neighbors, and $w_i$ is the negative average pairwise cosine similarity inside that neighborhood. Low-density neighborhoods receive high $w_i$, so tail samples are pulled harder toward their local mean; the loss is entirely unsupervised and, as written, has no repulsive term. The second load-bearing object is the self-guided labeling pipeline: long-tailed clustering estimates class sizes $\{\tilde{n}_c\}$, a learnable distribution $\pi$ is trained with a guided loss that includes a KL term to $\tilde{\pi}$, and Sinkhorn-Knopp over a large queue $Q'$ produces pseudo-labels $H(\pi,Q')$ used by the classification loss. A momentum encoder maintains the queues of probability vectors and representations that keep both objectives stable under small batch sizes.

What would settle it

Compute the mean pairwise cosine similarity of tail-class features on CIFAR-100-LT before and after the representation-balancing stage with T2=10 and K=5. If the similarities rise toward 1 while the reported test accuracy stays high, the balancing gain is likely an artifact of the evaluation protocol; if dispersion remains high and few-class accuracy still improves, the mechanism is real.

Watch

Extended reading notes

Core claim

The central claim is that the two failure modes of long-tailed generalized category discovery — biased pseudo-labels and head-dominated representations — are separable and each has a tractable fix. Self-guided labeling replaces the uniform target distribution of Sinkhorn-Knopp with a learnable distribution $\pi$ regularized toward an estimate $\tilde{\pi}$ from long-tailed clustering, and applies the pseudo-labeling over a large queue to reduce small-batch noise. Representation balancing defines a local density $w_i$ from pairwise cosine similarities in a sample's $K$-nearest-neighbor set and a local mean $\mu_i$, then minimizes $(1+w_i)(1-\cos(z_i,\mu_i))$, giving tail samples a stronger pull toward their neighborhood. The paper reports that these components together raise novel-class accuracy from 33.7% to 73.8% on CIFAR-100-LT relative to the ablated baseline, with ablations attributing +8.4 points to self-guided labeling and +7.6 to representation balancing.

Load-bearing premise

The second stage assumes that pulling every sample toward the mean of its fixed nearest neighbors for ten epochs, with no repulsive force, will sharpen tail clusters instead of collapsing all features into one point.

Editorial extensions

If this is right

  • On the paper's benchmarks, overall accuracy rises to 95.0% on CIFAR-10-LT, 74.4% on CIFAR-100-LT, 88.4% on ImageNet-100-LT, and 32.5% on Places-365-LT, beating the previous best by 3.9, 7.2, 4.7, and 2.6 points.
  • The two components are additive: ablations on CIFAR-100-LT attribute +8.4 points to self-guided labeling and +7.6 points to representation balancing over the baseline.
  • The gains hold across imbalance ratios from 20 to 150; at $\rho=150$ the method still beats the prior state-of-the-art by 7.1 points overall.
  • Because only labels for known classes are used, the same pipeline applies to any unlabeled collection with long-tailed known and novel classes, and the final backbone can be frozen and clustered for inference.

Reading between the lines

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

  • The neighborhood-density weight $w_i$ is a label-free proxy for tail-ness; attaching it to supervised long-tailed classifiers or to other self-supervised objectives would test whether the signal transfers beyond GCD, which the paper does not do.
  • The balancing loss has no repulsive term, so the reported gains may partly reflect shrinking feature norms rather than improved geometry; a feature-dispersion analysis before and after stage two would separate these explanations.
  • The learnable distribution $\pi$ can be read as a soft rebalancing of pseudo-labels, which connects the method to logit-adjustment and class-balanced losses in supervised long-tailed learning even though the paper does not draw that link.
  • The method assumes the total number of classes is known; removing that assumption is the natural next step toward fully unsupervised discovery in long-tailed data.
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 a two-stage framework for generalized category discovery (GCD) under long-tailed class distributions. The first stage combines MoCo-style contrastive representation learning with a self-guided labeling technique: a long-tailed clustering procedure estimates class frequencies, which guide a learnable target distribution pi; Sinkhorn-Knopp applied to a large queue produces pseudo-labels used to train a classification head. The second stage, called representation balancing, removes the classification head and trains only the backbone and projection head with a loss that pulls each representation toward the mean of its K nearest neighbors, weighted by a neighborhood-density term intended to emphasize tail classes. Experiments on CIFAR-10-LT, CIFAR-100-LT, ImageNet-100-LT, and Places-365-LT report state-of-the-art results (e.g., 95.0% overall on CIFAR-10-LT and 74.4% on CIFAR-100-LT), with ablations showing large contributions from both proposed components.

Significance. If the reported results hold, the paper makes a meaningful empirical contribution to a practical problem: GCD in long-tailed data is important for real-world deployment, and the claimed gains over BaCon (3.9–7.2 points overall) are substantial. The self-guided labeling idea, in which a learnable distribution is regularized toward a clustering-based estimate, is interesting and the neighborhood-density weighting for tail emphasis is a plausible mechanism. The paper is also transparent about hyperparameters and provides component ablations. However, the central mechanism of the second stage is not convincingly validated: the balancing loss is a pure attractive force with no described repulsive term, and the paper offers no feature-dispersion analysis to rule out collapse. The final clustering protocol is also underspecified. Because these gaps directly support the headline results, the manuscript is not ready for acceptance without additional evidence.

major comments (4)
  1. [3.3, Eq. (10)] The representation balancing stage trains only the backbone and projection head with Lbal = (1/B) sum_i (1 + w_i)(1 - cos(z_i, mu_i)), where mu_i is the mean of z_i and its K nearest neighbors. As written, this is a pure attractive objective: every sample is pulled toward the local mean, with no repulsive or contrastive term described in this stage. If the neighbors are not detached from the computation graph, the trivial fixed point of all z_i equal to the global mean is consistent with the loss; even with detached neighbors, repeated minimization of distance-to-local-mean can contract the feature cloud. This is load-bearing because Table 5(a) attributes a +7.6 point gain on CIFAR-100-LT to this stage, and the final evaluation clusters features from the frozen backbone. The paper provides no feature-dispersion measurements (average pairwise cosine similarity, effective rank, or singular value spectrum) before and after balancing, and no ablation with a stop-gradient or with an added contrastive term. Please add these measurements, clarify whether neighbors are detached, state whether z_i are L2-normalized, and describe whether any contrastive loss remains in this stage.
  2. [3.2, Eq. (2)] Equation (2) defines alpha as alpha = gamma - (1/M) sum_i sim(v'_i, hat{v}_c), which depends on the prototype index c, yet alpha is used as a scalar in the numerator of Phi_ic. As written, the formula is mathematically ill-posed: there is a different alpha_c for each cluster, and using one global alpha changes the sharpening of all clusters in a way that is not defined. Since this sharpening step feeds into Lcluster and the cluster-size estimates that guide the learnable distribution, this ambiguity is central to the self-guided labeling method. Please clarify whether alpha_c is intended and how it is computed in the implementation.
  3. [3.2, Eq. (4)] The learnable distribution pi is optimized through Lgud, whose first term depends on the Sinkhorn-Knopp pseudo-labels H(pi, Q'). The paper does not state how gradients are propagated through the Sinkhorn iterations, how the cluster-size parameters {n_c} are normalized to form a valid distribution, or whether pi is updated per batch or per epoch. Without these details, the 'learnable distribution' is not fully specified, and the self-referential loop among pi, the Sinkhorn pseudo-labels, and the model's own queue predictions cannot be properly assessed. Please provide the exact parameterization, the update rule, and the Sinkhorn regularization, and report sensitivity to the Sinkhorn temperature or iterations.
  4. [4.1 and 3.3] The final evaluation pipeline is underspecified. Section 3.3 ends with 'features, which are clustered to generate predictions,' but the paper does not state which clustering algorithm is used (e.g., semi-supervised k-means, k-means with C=|Yu|), how many iterations are run, how clusters are initialized, or whether labels from known classes are used in the final assignment. Since Table 2's headline numbers depend on this step, the exact inference procedure must be described. In addition, the main tables report no variance across random seeds; given that the claimed gains over BaCon range from 2.6 to 7.2 points, please report mean and standard deviation over at least three runs for the main comparisons.
minor comments (5)
  1. [3.3, Eq. (8)] The term 1[n,m] in Eq. (8) is undefined; it presumably indicates that the pair n=m is excluded from the double sum. Please define this notation explicitly.
  2. [4.2, Table 6] Hyperparameters gamma and lambda are not ablated; Table 6 only covers T1, beta, K, and T2. Please justify the fixed values of gamma and lambda, or add a sensitivity study.
  3. [4.2, Table 6(c)] In Table 6(c), K=6 yields a higher New accuracy (74.7) than the default K=5 (73.8), yet K=5 is chosen as the default. The stated reason of 'efficiency' is not quantified; please provide runtime or memory measurements to support this choice.
  4. [4.1] The implementation details for the second stage (optimizer, learning rate, number of epochs, data augmentation, and whether a momentum encoder is still used) are not reported. These details are needed for reproducibility and for assessing the collapse risk in the balancing stage.
  5. [Throughout] There are numerous typographical artifacts in the text (e.g., 'ef fective', 'di fferent', 'f eatures'). A careful copyedit is needed before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is held-out test accuracy, and the self-referential pseudo-labeling loop is standard self-training rather than a prediction reduced to its own inputs.

full rationale

The paper's central claim is test accuracy on disjoint balanced test sets, obtained by freezing the trained backbone, extracting features, clustering them, and matching clusters to ground truth with the Hungarian algorithm (Section 4.1, Table 2). The learnable distribution pi in the self-guided labeling loss (Eq. 4) is a training variable used to generate pseudo-labels for the classification loss; it is not a reported prediction, and the final evaluation does not use these pseudo-labels. Likewise, the representation balancing loss Lbal (Eq. 10) is a training objective, not a fitted quantity renamed as a result. The paper contains no self-citations: all cited methods (GCD, BaCon, SDCLR, etc.) are external prior work, and no uniqueness theorem or ansatz is imported from the author's own prior publications. The self-referential nature of pseudo-labeling and the potential feature-collapse risk of Lbal are correctness or robustness concerns, not circularity, because the benchmark against the external baselines is independently labeled and the ablations compare against other published losses. No step in the derivation chain is equivalent by construction to the input data or to a fitted parameter.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities, forces, or conserved quantities. Its novel components are parametric designs (a learnable distribution and a local-mean pull loss) whose assumptions are listed above.

free parameters (7)
  • Learnable distribution pi = Online learned vector of cluster sizes {n_c}, not reported
    The Sinkhorn target distribution pi with parameters {n_c} is optimized through L_gud (Eq. 4) and directly controls pseudo-label generation.
  • gamma = 2
    Controls the sharpening temperature alpha in the long-tailed clustering assignment (Eq. 2); chosen by the author.
  • beta = 400
    Scales the KL divergence term between the learnable distribution and the clustering estimate in L_gud (Eq. 4).
  • lambda = 0.35
    Blends supervised and unsupervised classification and representation losses (Eqs. 6 and 7).
  • T1 = 10
    Epoch interval for re-estimating the class distribution via clustering.
  • K = 5
    Number of nearest neighbors used in the representation balancing loss (Eq. 8).
  • T2 = 10
    Epoch interval for re-sampling neighborhoods in the balancing stage.
assumptions (5)
  • domain assumption The total number of classes in the combined dataset is known in advance.
    Stated in Section 3.1: 'Following BaCon [2], we assume that the number of classes in D is known in advance.' This sets the cluster count C and the classification head size.
  • domain assumption Pretrained ViT-B/16 features transfer well enough to initialize the backbone for long-tailed GCD.
    Section 4.1 uses a pre-trained ViT-B/16 to initialize the backbone; the paper supplies no feature-space analysis.
  • domain assumption The long-tailed clustering procedure (Eqs. 1-3) converges to cluster sizes that approximate the true class distribution.
    Section 3.2 derives the estimated distribution pi_tilde from cluster sizes after the cluster loss; no convergence or consistency guarantee is given.
  • ad hoc to paper The balanced loss Lbal preserves representation diversity and does not collapse the feature space.
    Section 3.3, Eq. 10 pulls each sample toward its local mean with no repulsive term; the paper presents no collapse analysis.
  • domain assumption Sinkhorn-Knopp with target distribution pi yields meaningful pseudo-labels for unlabeled data.
    Section 3.2 uses the output H(pi, Q') as training targets for cross-entropy; this is a standard but unproved assumption for the long-tailed setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Long-Tailed Learning for Generalized Category Discovery." pith.science (2026). https://pith.science/paper/CRJJEB5T

@misc{pith2026250606965,
  author       = {Pith},
  title        = {Pith review of: Long-Tailed Learning for Generalized Category Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRJJEB5T}},
  note         = {Machine review of arXiv:2506.06965}
}
read the original abstract

Generalized Category Discovery (GCD) utilizes labeled samples of known classes to discover novel classes in unlabeled samples. Existing methods show effective performance on artificial datasets with balanced distributions. However, real-world datasets are always imbalanced, significantly affecting the effectiveness of these methods. To solve this problem, we propose a novel framework that performs generalized category discovery in long-tailed distributions. We first present a self-guided labeling technique that uses a learnable distribution to generate pseudo-labels, resulting in less biased classifiers. We then introduce a representation balancing process to derive discriminative representations. By mining sample neighborhoods, this process encourages the model to focus more on tail classes. We conduct experiments on public datasets to demonstrate the effectiveness of the proposed framework. The results show that our model exceeds previous state-of-the-art methods.

Figures

Figures reproduced from arXiv: 2506.06965 by the authors.

Figure 1
Figure 1. (a) Self-guided labeling incorporates a learnable distribution that accurately reflects a long-tailed distribution during training; (b) Representation balancing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed framework. (a) Generalized category discovery includes a self-guided labeling process to produce pseudo-labels ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualization on the test set of CIFAR-10 [8]. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 47 canonical work pages

  1. [1]

    Masked siamese networks for label-efficient learning, in: European conference on computer vision, Springer

    Assran, M., Caron, M., Misra, I., Bojanowski, P., Bordes, F., Vincent, P., Joulin, A., Rabbat, M., Ballas, N., 2022. Masked siamese networks for label-efficient learning, in: European conference on computer vision, Springer. pp. 456–473

    work page 2022

  2. [2]

    Towards distribution-agnostic generalized cate- gory discovery

    Bai, J., Liu, Z., Wang, H., Chen, R., Mu, L., Li, X., Zhou, J.T., Feng, Y ., Wu, J., Hu, H., 2023. Towards distribution-agnostic generalized cate- gory discovery. Advances in Neural Information Processing Systems 36, 58625–58647

    work page 2023

  3. [3]

    Open- world semi-supervised learning, in: International Confer- ence on Learning Representations, pp

    Cao, K., Brbic, M., Leskovec, J., 2022. Open- world semi-supervised learning, in: International Confer- ence on Learning Representations, pp. 512–534. URL: https://openreview.net/forum?id=O-r8LOR-CCA

    work page 2022

  4. [4]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T., 2019. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems 32

    work page 2019

  5. [5]

    Emerging properties in self-supervised vision trans- formers, in: Proceedings of the IEEE /CVF international conference on computer vision, pp

    Caron, M., Touvron, H., Misra, I., J ´egou, H., Mairal, J., Bojanowski, P., Joulin, A., 2021. Emerging properties in self-supervised vision trans- formers, in: Proceedings of the IEEE /CVF international conference on computer vision, pp. 9650–9660

    2021

  6. [6]

    Smote: synthetic minority over-sampling technique

    Chawla, N.V ., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P., 2002. Smote: synthetic minority over-sampling technique. Journal of artificial intelli- gence research 16, 321–357

    work page 2002

  7. [7]

    Chen, J., Su, B., 2023. Transfer knowledge from head to tail: Uncer- tainty calibration under long-tailed distribution, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 19978–19987

    work page 2023

  8. [8]

    Class- balanced loss based on e ffective number of samples, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Cui, Y ., Jia, M., Lin, T.Y ., Song, Y ., Belongie, S., 2019. Class- balanced loss based on e ffective number of samples, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9268–9277

    work page 2019

Show all 50 references
  1. [9]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Cuturi, M., 2013. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems 26

    2013

  2. [10]

    LPT: Long-tailed prompt tuning for image classification, in: The Eleventh International Conference on Learning Representations, pp

    Dong, B., Zhou, P., Y AN, S., Zuo, W., 2023. LPT: Long-tailed prompt tuning for image classification, in: The Eleventh International Conference on Learning Representations, pp. 41229–41248. URL: https://openreview.net/forum?id=8pOVAeo8ie

    2023

  3. [11]

    Fini, E., Sangineto, E., Lathuili `ere, S., Zhong, Z., Nabi, M., Ricci, E.,

  4. [12]

    Learning to discover and detect objects

    Fomenko, V ., Elezi, I., Ramanan, D., Leal-Taix ´e, L., Osep, A., 2022. Learning to discover and detect objects. Advances in Neural Information Processing Systems 35, 8746–8759

    2022

  5. [13]

    Enhancing minority classes by mixing: An adaptative optimal transport approach for long-tailed classifi- cation

    Gao, J., Zhao, H., Li, Z., Guo, D., 2023. Enhancing minority classes by mixing: An adaptative optimal transport approach for long-tailed classifi- cation. Advances in Neural Information Processing Systems 36, 60329– 60348

    2023

  6. [14]

    Autonovel: Automatically discovering and learning novel visual cate- gories

    Han, K., Rebu ffi, S.A., Ehrhardt, S., Vedaldi, A., Zisserman, A., 2021. Autonovel: Automatically discovering and learning novel visual cate- gories. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 6767–6781

    2021

  7. [15]

    Momentum contrast for unsupervised visual representation learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    He, K., Fan, H., Wu, Y ., Xie, S., Girshick, R., 2020. Momentum contrast for unsupervised visual representation learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9729–9738

    2020

  8. [16]

    Learning to cluster in order to trans- fer across domains and tasks, in: International Conference on Learning Representations, pp

    Hsu, Y .C., Lv, Z., Kira, Z., 2018. Learning to cluster in order to trans- fer across domains and tasks, in: International Conference on Learning Representations, pp. 2567–2581

    2018

  9. [17]

    Learning discrete representations via information maximizing self- augmented training, in: International conference on machine learning, PMLR

    Hu, W., Miyato, T., Tokui, S., Matsumoto, E., Sugiyama, M., 2017. Learning discrete representations via information maximizing self- augmented training, in: International conference on machine learning, PMLR. pp. 1558–1567

    2017

  10. [18]

    Self-damaging contrastive learning, in: International Conference on Machine Learning, PMLR

    Jiang, Z., Chen, T., Mortazavi, B.J., Wang, Z., 2021. Self-damaging contrastive learning, in: International Conference on Machine Learning, PMLR. pp. 4927–4939

    2021

  11. [19]

    Jin, Y ., Li, M., Lu, Y ., Cheung, Y .m., Wang, H., 2023. Long-tailed visual recognition via self-heterogeneous integration with knowledge excava- tion, in: Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp. 23695–23704

    2023

  12. [20]

    Decoupling representation and classifier for long-tailed recog- nition, in: Eighth International Conference on Learning Representations (ICLR), pp

    Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y ., 2020. Decoupling representation and classifier for long-tailed recog- nition, in: Eighth International Conference on Learning Representations (ICLR), pp. 5319–5328

    2020

  13. [21]

    Distri- bution aligning refinery of pseudo-label for imbalanced semi-supervised learning

    Kim, J., Hur, Y ., Park, S., Yang, E., Hwang, S.J., Shin, J., 2020a. Distri- bution aligning refinery of pseudo-label for imbalanced semi-supervised learning. Advances in neural information processing systems 33, 14567– 14579

  14. [22]

    M2m: Imbalanced classification via major-to-minor translation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Kim, J., Jeong, J., Shin, J., 2020b. M2m: Imbalanced classification via major-to-minor translation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 13896–13905

  15. [23]

    Temperature schedules for self-supervised contrastive methods on long-tail data, in: ICLR, pp

    Kukleva, A., B ¨ohle, M., Schiele, B., Kuehne, H., Rupprecht, C., 2023. Temperature schedules for self-supervised contrastive methods on long-tail data, in: ICLR, pp. 1683–1692. URL: https://openreview.net/forum?id=ejHUr4nfHhD

    2023

  16. [24]

    Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning

    Lee, H., Shin, S., Kim, H., 2021. Abc: Auxiliary balanced classifier for class-imbalanced semi-supervised learning. Advances in Neural Informa- tion Processing Systems 34, 7082–7094

    2021

  17. [25]

    Long-tailed visual recognition via gaussian clouded logit adjustment, in: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pp

    Li, M., Cheung, Y .m., Lu, Y ., 2022. Long-tailed visual recognition via gaussian clouded logit adjustment, in: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pp. 6929–6938

    2022

  18. [26]

    Im- proving visual prompt tuning by gaussian neighborhood minimization for long-tailed visual recognition

    Li, M., Liu, Y ., Lu, Y ., Zhang, Y ., Cheung, Y .m., Huang, H., 2024a. Im- proving visual prompt tuning by gaussian neighborhood minimization for long-tailed visual recognition. Advances in Neural Information Process- ing Systems 37, 103985–104009

  19. [27]

    Feature fusion from head to tail for long-tailed visual recognition, in: Proceedings of the AAAI conference on artificial intelligence, pp

    Li, M., Zhikai, H., Lu, Y ., Lan, W., Cheung, Y .m., Huang, H., 2024b. Feature fusion from head to tail for long-tailed visual recognition, in: Proceedings of the AAAI conference on artificial intelligence, pp. 13581– 13589

  20. [28]

    Conmix: Contrastive mixup at representa- tion level for long-tailed deep clustering, in: The Thirteenth Interna- tional Conference on Learning Representations, pp

    Li, Z., Jia, Y ., 2025. Conmix: Contrastive mixup at representa- tion level for long-tailed deep clustering, in: The Thirteenth Interna- tional Conference on Learning Representations, pp. 653–667. URL: https://openreview.net/forum?id=3lH8WT0fhu

    2025

  21. [29]

    Exploratory undersampling for class- imbalance learning

    Liu, X.Y ., Wu, J., Zhou, Z.H., 2008. Exploratory undersampling for class- imbalance learning. IEEE Transactions on Systems, Man, and Cybernet- ics, Part B (Cybernetics) 39, 539–550

    2008

  22. [30]

    Large- scale long-tailed recognition in an open world, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X., 2019. Large- scale long-tailed recognition in an open world, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2537–2546

    2019

  23. [31]

    Protogcd: Uni- fied and unbiased prototype learning for generalized category discov- ery

    Ma, S., Zhu, F., Zhang, X.Y ., Liu, C.L., 2025. Protogcd: Uni- fied and unbiased prototype learning for generalized category discov- ery. IEEE Transactions on Pattern Analysis and Machine Intelligence , 1–17doi:10.1109/TPAMI.2025.3557502

    2025

  24. [32]

    One million scenes for autonomous driving: Once dataset

    Mao, J., Niu, M., Jiang, C., Liang, H., Liang, X., Li, Y ., Ye, C., Zhang, W., Li, Z., Yu, J., et al., 2021. One million scenes for autonomous driving: Once dataset. NeurIPS

    2021

  25. [33]

    Fighting class imbalance with contrastive learning, in: International conference on medical image computing and computer-assisted intervention, Springer

    Marrakchi, Y ., Makansi, O., Brox, T., 2021. Fighting class imbalance with contrastive learning, in: International conference on medical image computing and computer-assisted intervention, Springer. pp. 466–476

    2021

  26. [34]

    Park, S., Hong, Y ., Heo, B., Yun, S., Choi, J.Y ., 2022. The majority can help the minority: Context-rich minority oversampling for long-tailed classification, in: Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, pp. 6887–6896

    2022

  27. [35]

    Towards realistic semi- supervised learning, in: European Conference on Computer Vision, Springer

    Rizve, M.N., Kardan, N., Shah, M., 2022. Towards realistic semi- supervised learning, in: European Conference on Computer Vision, Springer. pp. 437–455

    2022

  28. [36]

    Long- tail learning with foundation model: Heavy fine-tuning hurts, in: Interna- tional Conference on Machine Learning, PMLR

    Shi, J.X., Wei, T., Zhou, Z., Shao, J.J., Han, X.Y ., Li, Y .F., 2024. Long- tail learning with foundation model: Heavy fine-tuning hurts, in: Interna- tional Conference on Machine Learning, PMLR. pp. 45014–45039

    2024

  29. [37]

    Opencon: Open-world contrastive learning, in: Transactions on Machine Learning Research, pp

    Sun, Y ., Li, Y ., 2023. Opencon: Open-world contrastive learning, in: Transactions on Machine Learning Research, pp. 5182–5194. URL: https://openreview.net/forum?id=2wWJxtpFer

    2023

  30. [38]

    Equalization loss for long-tailed object recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J., 2020. Equalization loss for long-tailed object recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11662–11671. 9

    2020

  31. [39]

    Contrastive multiview coding, in: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, Springer

    Tian, Y ., Krishnan, D., Isola, P., 2020. Contrastive multiview coding, in: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, Springer. pp. 776– 794

    2020

  32. [40]

    The inaturalist species classi- fication and detection dataset, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Van Horn, G., Mac Aodha, O., Song, Y ., Cui, Y ., Sun, C., Shepard, A., Adam, H., Perona, P., Belongie, S., 2018. The inaturalist species classi- fication and detection dataset, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 8769–8778

    2018

  33. [41]

    Generalized category discovery, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp

    Vaze, S., Han, K., Vedaldi, A., Zisserman, A., 2022. Generalized category discovery, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp. 7492–7501

    2022

  34. [42]

    Adap- tive class suppression loss for long-tail object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Wang, T., Zhu, Y ., Zhao, C., Zeng, W., Wang, J., Tang, M., 2021. Adap- tive class suppression loss for long-tail object detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3103–3112

    2021

  35. [43]

    Parametric classification for generalized category discovery: A baseline study, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, pp

    Wen, X., Zhao, B., Qi, X., 2023. Parametric classification for generalized category discovery: A baseline study, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, pp. 16590–16600

    2023

  36. [44]

    Divide and con- quer: Compositional experts for generalized novel class discovery, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pp

    Yang, M., Zhu, Y ., Yu, J., Wu, A., Deng, C., 2022. Divide and con- quer: Compositional experts for generalized novel class discovery, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pp. 14268–14277

    2022

  37. [45]

    Novel class discovery for long-tailed recognition

    Zhang, C., Xu, R., He, X., 2023a. Novel class discovery for long-tailed recognition. Transactions on Machine Learning Research

  38. [46]

    Deep long-tailed learning: A survey

    Zhang, Y ., Kang, B., Hooi, B., Yan, S., Feng, J., 2023b. Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence 45, 10795–10816

  39. [47]

    Improving calibration for long- tailed recognition, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pp

    Zhong, Z., Cui, J., Liu, S., Jia, J., 2021. Improving calibration for long- tailed recognition, in: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pp. 16489–16498

    2021

  40. [48]

    Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Zhou, B., Cui, Q., Wei, X.S., Chen, Z.M., 2020. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9719–9728

    2020

  41. [49]

    Places: A 10 million image database for scene recognition

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A., 2017. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40, 1452–1464. 10

    2017

  42. [2021]

    9284– 9292

    A unified objective for novel class discovery, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9284– 9292

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.