Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Mixture of Balanced Information Bottlenecks for Long-Tailed Visual Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper argues that a loss combining balanced softmax cross-entropy with variational self-distillation implicitly optimizes the information bottleneck objective, and applying it at three network depths rivals mixture-of-experts accuracy o

desk verdict Solid empirical recipe for long-tailed recognition, but the information-bottleneck story doesn't survive contact with the math; the method may still work, but the paper's main claim needs reworking. read the letter →

arxiv 2509.01804 v1 pith:PA2UMU32 submitted 2025-09-01 cs.CV cs.ITmath.IT

classification cs.CVcs.ITmath.IT
keywords long-tailedrecognitioninformationbottleneckself-distillationbalancedsoftmaxlosslogitadjustmentrepresentationlearningclassimbalancemixture-of-experts
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

This paper tries to establish that long-tailed visual recognition—where a few head classes dominate training data and most tail classes have few samples—can be solved end-to-end by a loss that rebalances classes while implicitly optimizing the information bottleneck objective. The central object is the BIB loss, which adds a variational self-distillation term to two balanced softmax cross-entropy losses; the paper argues this combination maximizes label-relevant information and compresses label-irrelevant information without requiring explicit mutual information estimates. A structural extension, MBIB, applies the same loss at three network depths, exploiting the data processing inequality to harvest label information that shallower layers retain. If the argument holds, single-expert models trained in one pass can match or approach mixture-of-experts accuracy on long-tailed benchmarks at roughly half the compute.

What carries the argument

The pivotal identity is I(v;z) = I(z;y) + I(v;z|y), which splits the bottleneck objective into three sub-objectives. The load-bearing mechanism is the variational self-distillation loss, Loss3 = E[D_KL(q(y|v)||q(y|z))]: minimizing the KL divergence between the classifier outputs at v and z makes H(y|v) ≈ H(y|z), which the paper takes to force I(z;y) ≈ I(v;y) without explicit mutual-information estimation. Loss1 and Loss2 are balanced softmax cross-entropy losses that add log n_i to the logits and weight classes inversely to frequency, rebalancing the long-tailed label distribution. MBIB's structural argument is the data processing inequality I(v3;y) ≤ I(v2;y) ≤ I(v1;y), showing shallower rep

What would settle it

A direct test: train BIB on a synthetic long-tailed dataset whose label depends only on a few input dimensions, then measure I(v;z|y) and |I(v;y) − I(z;y)| as β grows. If the self-distillation loss can be driven near zero while z still retains label-independent noise from x (I(v;z|y) stays large), the claimed equivalence fails. Concretely, check whether increasing β continues to compress z's label-irrelevant dimensions after the KL term has already converged.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that minimizing LossBIB(v,z) = Loss1 + Loss2 + β·Loss3 implicitly solves the information bottleneck problem under a long-tailed label distribution. Under the assumption that encoder output v preserves all label-relevant information from x, the IB objective decomposes via I(v;z) = I(z;y) + I(v;z|y) into three sub-objectives: maximize I(v;y), maximize I(z;y), and minimize |I(v;y) − I(z;y)|. Loss1 and Loss2 are balanced softmax cross-entropy losses that maximize the first two terms; Loss3 is a variational self-distillation loss minimizing D_KL[q(y|v)||q(y|z)], which drives I(z;y) toward I(v;y). MBIB sums BIB losses at three depths, a·LossBIB(v1,z)

Load-bearing premise

The whole argument rides on the claim that driving two classifiers' predictions together is equivalent to driving the mutual information between their representations and the label together; if that equivalence needs conditions the paper does not state, Loss3 is not a valid information bottleneck proxy and what remains is an empirical loss combination.

Editorial extensions

If this is right

  • Long-tailed recognition can be trained end-to-end in one stage: representation and classifier are learned jointly from a single loss, avoiding the representation/classifier decoupling used by two-stage methods.
  • Single-expert models become a practical alternative to mixture-of-experts: the paper reports MBIB at 486K parameters and 27.93 GFLOPs versus 784K/40.69 G for SADE and 1018K/50.98 G for RIDE(4E), with competitive or better accuracy on iNaturalist 2018.
  • Both rebalancing ingredients matter: ablations show removing logit adjustment costs 0.9% overall accuracy and removing class rebalancing costs 2.2%, so the balanced cross-entropy design is load-bearing, not decorative.
  • The self-distillation term (β > 0) is what makes the loss a bottleneck: ablation results show the balanced terms alone (β = 0) underperform BIB, consistent with the claim that compressing label-irrelevant information helps long-tailed recognition.

Reading between the lines

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

  • The paper's information-theoretic reading of Loss3 is one of several possible descriptions: the same objective can be seen as standard teacher-student self-distillation between two heads of one network. If the equivalence between KL matching and mutual-information matching fails in some regime (e.g., when v is a poor statistic of x), the method's empirical gains might still hold, but the IB interp
  • The ablation on the number of observations v (2-v through 6-v) suggests a general diminishing-returns law for layer-wise auxiliary supervision: each added bottleneck helps until the optimization cost outweighs the extra label information. This is a testable design principle for other long-tailed architectures.
  • The Appendix G result that adding MBIB to the SADE expert framework improves accuracy by 1.6% hints that the bottleneck mechanism and expert routing are complementary—one could combine them systematically, e.g., one BIB per expert with expert-specific weights.
  • The class-dependent temperature Ti = (n_max/n_i)^γ couples distillation strength to class frequency; tuning γ per class rather than globally might extract more tail-class signal, an extension the paper does not explore.
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

3 major / 5 minor

Summary. The paper proposes Balanced Information Bottleneck (BIB) and Mixture of Balanced Information Bottlenecks (MBIB) for long-tailed visual recognition. BIB combines two balanced softmax cross-entropy losses (Loss1, Loss2) with a variational self-distillation loss (Loss3) and claims this implicitly optimizes the information bottleneck objective. MBIB extends BIB by summing BIB losses at three network depths, using intermediate representations to preserve label-related information. Experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 report accuracy competitive with or better than many one-stage, two-stage, and mixture-of-experts baselines, along with an efficiency comparison.

Significance. If the theoretical claim were sound, BIB/MBIB would offer a useful single-expert alternative to MoE-style long-tailed methods, with lower compute and end-to-end training. The paper has empirical strengths: evaluations span three standard datasets and multiple backbones; ablations address loss components, re-balancing choices, and number of observations; efficiency is compared with MoE baselines; and Python code for the BIB loss is included in Appendix J. These strengths support the practical potential of the method. However, the information-bottleneck interpretation is the paper's central contribution and is not established: Theorem 3.2 contains a load-bearing equivalence that is false under the stated Markov structure, and the proposed loss contains no actual compression term. The empirical method may still be valuable, but the current framing overstates the theoretical guarantee.

major comments (3)
  1. [Sec. 3.2, Theorem 3.2, Eq. (3)] The proof claims that minimizing I(v;z|y) is equivalent to minimizing |I(v;y)-I(z;y)|, citing Tian et al. Under the natural Markov chain y→v→z, we have I(z;y|v)=0, so I(v;y)-I(z;y)=I(v;y|z), not I(v;z|y). These conditional mutual informations are not generally equal. For example, let Z be a noisy copy of Y and let V=(Z,N) with N independent of Y given Z. Then Z is sufficient for Y, giving I(v;y|z)=0 and |I(v;y)-I(z;y)|=0, while I(v;z|y)=H(z|y)>0. Thus Loss3, which matches q(y|v) and q(y|z), is not a valid proxy for the compression term I(v;z|y). Eq. (9) therefore contains no term that minimizes the bottleneck quantity I(v;z|y) or I(v;z), and the central claim that LossBIB implicitly optimizes the IB objective is unsupported. The same issue propagates to Eq. (11) and the MBIB motivation.
  2. [Sec. 3.2, Eq. (8) and implementation] The VSD loss actually used in experiments is not the KL divergence in Eq. (8) when gamma>0. Immediately after Eq. (8), class-dependent temperatures T_i=(n_max/n_i)^gamma are inserted into q(y|v) and q(y|z). For ImageNet-LT and iNaturalist, gamma=0.5, so the loss is a temperature-scaled, class-weighted KL. No information-theoretic derivation is given for this scaling, and it changes the conditional entropy and mutual-information relationships on which Theorem 3.2 relies. Either the theorem and loss need to be formulated for the temperature-scaled distributions, or the experiments need to be run with gamma=0 so that Eq. (8) matches the implementation.
  3. [Sec. 4.2, Figures 3 and 4] The hyperparameters beta, a, and b are chosen by looking at overall accuracy on the test sets (e.g., 'we take beta in {0,1,2,3,4,5} according to different datasets' and 'we selected a=0.1 and b=0.3 within these satisfactory configurations'). No validation split or selection protocol is described. If the reported numbers were selected using test-set accuracy, the state-of-the-art comparisons in Tables 1-3 are at risk of selection bias. The paper should either use a held-out validation set or explicitly report the selection procedure and its influence on the reported gains.
minor comments (5)
  1. [Sec. 1 / References] In the introduction, 'BBN (Wang et al., 2020), RIDE (Zhou et al., 2020)' appears to swap the citations: BBN is Zhou et al. (2020) and RIDE is Wang et al. (2020).
  2. [Eq. (8)] The notation 'q(y|v)[log q(y|v)-log q(y|z)]' is ambiguous; the expectation should be written as E_{q(y|v)}[log q(y|v)-log q(y|z)].
  3. [Appendix C] The text says 'We assume v can retain all the information in x, but the information will still be lost from x to v.' This directly contradicts Assumption 3.1, which states I(v;y)=I(x;y). Please reconcile the assumption with the acknowledged information loss.
  4. [Table 6] The efficiency comparison reports Params and FLOPs but does not specify the backbone or training/inference configuration for each method. Since RIDE and SADE use different backbones in the main experiments, clarify that the comparison is apples-to-apples.
  5. [Appendix J] The code hard-codes beta=5 and label_weighting=0.1 rather than taking them as arguments. This makes it hard to reproduce the hyperparameter search described in Sec. 4.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical, and the information-theoretic framing, while relying on an external citation, does not reduce to its inputs by construction.

full rationale

The paper's headline claims are empirical: BIB and MBIB improve accuracy on long-tailed benchmarks relative to baselines. The loss (Eq. 9) is a weighted sum of two balanced softmax cross-entropy losses and a KL-divergence self-distillation term; no fitted parameter is later presented as a prediction of that same parameter. The information-theoretic justification in Theorem 3.2 rests on an equivalence attributed to Tian et al. (2021) — an external reference, not a self-citation. Whether that equivalence is correct is a question of mathematical validity, not circularity: the paper does not define the IB objective as its own loss, nor does any equation reduce the stated 'prediction' to the training inputs by construction. The MBIB objective (Eq. 11) is a linear combination of BIB losses, again an empirical design. The absence of any self-citation chain and the presence of external benchmarks (CIFAR-100-LT, ImageNet-LT, iNaturalist 2018) with reported test accuracies further support a non-circular finding. The main caveat is that the cited equivalence is not proved in the paper and may be false; that is a correctness/rigor concern, not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. It does introduce five fitted hyperparameters and relies on two non-trivial assumptions: the sufficiency of v, which the paper itself concedes is violated, and the Tian et al. equivalence, which is unproved in this paper. These assumptions are load-bearing for the information-theoretic interpretation of the loss.

free parameters (5)
  • beta = selected from {0,1,2,3,4,5} per dataset
    Controls the weight of the variational self-distillation loss; chosen by grid search on test accuracy (Section 4.2, Figure 3).
  • a = 0.1
    Weight for LossBIB(v1,z) in MBIB; tuned on CIFAR-100-LT accuracy heatmap and fixed for all datasets (Section 4.2, Figure 4).
  • b = 0.3
    Weight for LossBIB(v2,z) in MBIB; tuned on CIFAR-100-LT accuracy heatmap and fixed for all datasets (Section 4.2, Figure 4).
  • m = 0.1
    Exponent for class re-balancing weights in Eq. (7); copied from prior work (Du et al., 2023) and affects Loss1 and Loss2.
  • gamma = 0 for CIFAR-100-LT, 0.5 for ImageNet-LT and iNaturalist
    Exponent for class-dependent self-distillation temperature; selected per dataset (Section 4.2).
assumptions (4)
  • ad hoc to paper Assumption 3.1: the observation v extracted by the encoder retains all label-relevant information, I(v;y)=I(x;y).
    Stated in Section 3.2 to make the IB objective tractable, but the paper itself later argues in Theorem 3.3 that this assumption fails in practice due to the data processing inequality.
  • domain assumption Minimizing I(v;z|y) is equivalent to minimizing |I(v;y)-I(z;y)|, and minimizing KL(q(y|v)||q(y|z)) makes H(y|v) approximate H(y|z).
    Borrowed from Tian et al. (2021) and used in Theorem 3.2 and Eq. (8). The paper does not prove or state conditions for this equivalence, and Eq. (3) alone does not establish it.
  • standard math The variables follow a Markov chain y -> x -> v -> z, and variational distributions q(y|v), q(y|z), q(z|x), r(z) can approximate the true distributions.
    Standard information bottleneck assumption, used implicitly in Section 3.2 and Eq. (2) following Alemi et al. (2017).
  • standard math Data processing inequality: I(v3;y) <= I(v2;y) <= I(v1;y) for intermediate representations of a CNN.
    Invoked as Theorem 3.3 and used to justify multi-layer BIB; it is a standard consequence of DPI for Markov chains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixture of Balanced Information Bottlenecks for Long-Tailed Visual Recognition." pith.science (2026). https://pith.science/paper/PA2UMU32

@misc{pith2026250901804,
  author       = {Pith},
  title        = {Pith review of: Mixture of Balanced Information Bottlenecks for Long-Tailed Visual Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PA2UMU32}},
  note         = {Machine review of arXiv:2509.01804}
}
read the original abstract

Deep neural networks (DNNs) have achieved significant success in various applications with large-scale and balanced data. However, data in real-world visual recognition are usually long-tailed, bringing challenges to efficient training and deployment of DNNs. Information bottleneck (IB) is an elegant approach for representation learning. In this paper, we propose a balanced information bottleneck (BIB) approach, in which loss function re-balancing and self-distillation techniques are integrated into the original IB network. BIB is thus capable of learning a sufficient representation with essential label-related information fully preserved for long-tailed visual recognition. To further enhance the representation learning capability, we also propose a novel structure of mixture of multiple balanced information bottlenecks (MBIB), where different BIBs are responsible for combining knowledge from different network layers. MBIB facilitates an end-to-end learning strategy that trains representation and classification simultaneously from an information theory perspective. We conduct experiments on commonly used long-tailed datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist 2018. Both BIB and MBIB reach state-of-the-art performance for long-tailed visual recognition.

Figures

Figures reproduced from arXiv: 2509.01804 by the authors.

Figure 1
Figure 1. Considering CNNs’ powerful feature extraction capability, we introduce the following assumption [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 1
Figure 1. The network structure of BIB. FC means the Fully Connected layer and GAP means Global [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The network structure of MBIB. CNN1, CNN2 and CNN3 are different parts of the CNN network. FC means the Fully Connected layer, and GAP means Global Average Pooling. h(v1; θ), u(v2; θ), f(v3; θ) and g(z; θ) are classifiers. The output is the mean of the outputs from f(v3; θ) and g(z; θ). where there are two balanced cross entropy losses and one variational self-distillation loss. From the perspective of self-distilla… view at source ↗
Figures from the paper (8 more)
Figure 3
Figure 3. Figure 3: The impact of different β on the overall accuracy of CIFAR-100-LT (we fixed a = 0.1 and b = 0.3). The solid lines show the results of MBIB and the dashed lines correspond to BIB [PITH_FULL_IMAGE:figures/full_fig_p011_3.png]
Figure 4
Figure 4. Figure 4: The impact of different a and b on the overall MBIB accuracy of CIFAR-100-LT (we fixed β = 5). inter-class distance (DInter), and the ratio (ρ) between them. DIntra, DInter and ρ are calculated as follows: DIntra = 1 K X K i=1 1 |Ri | 2 X rj ,rk∈Ri ∥ rj − rk∥2 , (15) D…
Figure 5
Figure 5. Figure 5: Two alternative multi-BIB structures. BIB(v2, v3), and BIB(v3, z)), as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Ablation results of loss function. The Balanced option is [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: The mean Positive Posterior probability on ImageNet-LT with ResNet10. (a) The mean Positive [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Visualization analysis. The t-SNE is used to visualize the test set feature space on CIFAR-100-LT [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Overall accuracy of the three MBIB networks on CIFAR-100-LT (IF=100). [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Python code of BIB Loss. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 17 canonical work pages

  1. [1]

    Alemi, Ian S

    Alexander A. Alemi, Ian S. Fischer, Joshua V. Dillon, and Kevin P. Murphy. Deep variational information bottleneck. ArXiv, abs/1612.00410,

  2. [5]

    Drop-Bottleneck: Learning Discrete Compressed Representation for Noise-Robust Exploration

    Jaekyeom Kim, Minjung Kim, Dongyeon Woo, and Gunhee Kim. Drop-bottleneck: Learning discrete com- pressed representation for noise-robust exploration.ArXiv, abs/2103.12300,

  3. [7]

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2532–2541,

  4. [9]

    Long-tail learning via logit adjustment.ArXiv, abs/2007.07314,

    Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment.ArXiv, abs/2007.07314,

  5. [10]

    Decoupled Training for Long-Tailed Classification With Stochastic Representations

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

  6. [12]

    The information bottleneck method

    N Tishby. The information bottleneck method. InProc. 37th Annual Allerton Conference on Communica- tions, Control and Computing, 1999, pp. 368–377,

  7. [15]

    Margin Calibration for Long-Tailed Visual Recognition

    Yidong Wang, Bowen Zhang, Wenxin Hou, Zhen Wu, Jindong Wang, and Takahiro Shinozaki. Margin calibration for long-tailed visual recognition.ArXiv, abs/2112.07225, 2021d. Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning.ArXiv, abs/2006.07529,

  8. [16]

    Distribution alignment: A unified framework for long-tail visual recognition.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2361–2370,

Show all 21 references
  1. [17]

    Deep long-tailed learning: A survey

    Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023a. Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A ...

  2. [18]

    Rebalanced siamese contrastive mining for long-tailed recognition.ArXiv, abs/2203.11506,

    Zhisheng Zhong, Jiequan Cui, Eric Lo, Zeming Li, Jian Sun, and Jiaya Jia. Rebalanced siamese contrastive mining for long-tailed recognition.ArXiv, abs/2203.11506,

  3. [19]

    However, when the labels are long-tailed, we need to re-balance them

    A Proof of the Re-Balance Technique According to VIB,−I(v,y ) and−I(z,y ) are bounded asEp(v,y)−logq(y|v) and Ep(z,y)−logq(y|z). However, when the labels are long-tailed, we need to re-balance them. Our purpose is to train an end-to-end model, that is, the output of the model ...

  4. [20]

    To address these issues, logit adjustment methods have been proposed (Menon et al., 2021)

    have highlighted their limitations, including the lack of Fisher consistency. To address these issues, logit adjustment methods have been proposed (Menon et al., 2021). Furthermore, these methods can be combined, as demonstrated in Zhang et al. (2023a). As a result, we incorpo...

  5. [21]

    MBIB w/o logits adjustment means removing logits adjustment in MBIB

    MBIB w/o class re-balancing meansm = 0 and the weight in Eq.7 is 1 for all classes. MBIB w/o logits adjustment means removing logits adjustment in MBIB. MBIB w/o re-balancing means not using any re-balancing techniques in MBIB. The experimental results reveal that removing eit...

  6. [1999]

    Deep learning and the information bottleneck principle.2015 IEEE Information Theory Workshop (ITW), pp

    Naftali Tishby and Noga Zaslavsky. Deep learning and the information bottleneck principle.2015 IEEE Information Theory Workshop (ITW), pp. 1–5,

  7. [2006]

    Farewell to mutual information: Variational distillation for cross-modal person re-identification.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Xudong Tian, Zhizhong Zhang, Shaohui Lin, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Farewell to mutual information: Variational distillation for cross-modal person re-identification.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1522–1531,

  8. [2008]

    Label-aware distribution calibration for long-tailed classification.ArXiv, abs/2111.04901, 2021a

    Chaozheng Wang, Shuzheng Gao, Cuiyun Gao, Pengyun Wang, Wenjie Pei, Lujia Pan, and Zenglin Xu. Label-aware distribution calibration for long-tailed classification.ArXiv, abs/2111.04901, 2021a. Jing Wang, Yuanjie Zheng, Jingqi Song, and Sujuan Hou. Cross-view representation lea...

  9. [2019]

    Significance-aware information bottleneck for domain adaptive semantic segmentation.2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Yawei Luo, Ping Liu, Tao Guan, Junqing Yu, and Yi Yang. Significance-aware information bottleneck for domain adaptive semantic segmentation.2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6777–6786,

  10. [2020]

    HaoChen, Adrien Gaidon, and Tengyu Ma

    Hong Liu, Jeff Z. HaoChen, Adrien Gaidon, and Tengyu Ma. Self-supervised learning is more robust to dataset imbalance. ArXiv, abs/2110.05025,

  11. [2021]

    Disentangling label distribution for long-tailed visual recognition.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6622–6632,

  12. [2022]

    Belongie

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge J. Belongie. Class-balanced loss based on effective number of samples.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9260–9269,

  13. [2023]

    Botvinick, H

    Anirudh Goyal, Riashat Islam, Daniel Strouse, Zafarali Ahmed, Matthew M. Botvinick, H. Larochelle, Sergey Levine, and Yoshua Bengio. Infobot: Transfer and exploration via the information bottleneck.ArXiv, abs/1901.10902,

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.