Pith. sign in

REVIEW 3 major objections 4 minor 54 references

Mitigating Spurious Correlations with Causal Logit Perturbation

T0 review · 3 major / 4 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 The paper claims that training classifiers with sample-wise, meta-learned logit perturbations that encode human causal knowledge breaks spurious background-label correlations, and shows state-of-the-art accuracy on four biased-learning…

desk verdict Solid meta-learned logit perturbation with consistent gains, but the causal-intervention claim is asserted, not derived; deserves review with a request to fix the framing. read the letter →

arxiv 2505.15246 v1 pith:EIQWE4WJ submitted 2025-05-21 cs.LG

Xiaoling Zhou , Wei Ye , Rui Xie , Shikun Zhang This is my paper · ORCID
classification cs.LG
keywords spuriouscorrelationslogitperturbationcausalinferencemeta-learningcounterfactualaugmentationlong-taillearningnoisylabelssubpopulationshift
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 sets out to show that a classifier trained on ordinary biased data can be made to ignore spurious background-label correlations by perturbing its logits during training. The proposed CLP framework learns, through a small meta-learning loop, to generate a per-sample logit perturbation that acts as an intervention on the non-causal component of the model's output, closing the backdoor path from background confounders to the label. A reader should care because this offers a single recipe, clean metadata plus foreground masks plus counterfactual and factual augmentations, that the paper reports as state of the art in four distinct biased-learning settings: long-tail learning, noisy labels, subpopulation shifts, and generalized long-tail learning.

What carries the argument

The load-bearing object is the sample-wise causal logit perturbation $\delta_x$, produced by a two-layer MLP perturbation network from ten training characteristics of each sample, including loss, margin, gradient norm, class proportion, and relative loss and margin. The network is optimized by a bi-level meta-learning loop: an inner step updates the classifier on perturbed logits, and an outer step updates the perturbation network on the augmented metadata using a classification loss plus a saliency regularization $R^{Sal}$ that penalizes gradients on non-causal pixels. The metadata augmentation injects causal knowledge: counterfactual infilling $\Phi_{cf}$ replaces the foreground so the label flips to non-$y$, while factual infilling $\Phi_f$ replaces the background so the label stays $y$. Together these components let the perturbation network learn to push each sample's logits in a direction that reduces reliance on background information.

What would settle it

Train CLP on a benchmark with known ground-truth masks, once with the true masks and once with corrupted masks (for example, bounding boxes shifted so they include background, or shrunk so they cut off the object). If worst-group accuracy stays nearly the same in both runs, then accurate causal information is not what drives the reported gains; if accuracy collapses, the foreground-mask assumption is the load-bearing link.

Watch

Extended reading notes

Core claim

The paper's central claim is that logit perturbation can implement causal intervention: instead of reweighting data or hand-designing losses, CLP trains a perturbation network to output a perturbation $\delta_x$ such that training on the perturbed logits $u_x + \delta_x$ is equivalent to intervening on the extrinsic component $u_z$ of the logit vector. Under the paper's causal graph, a confounder $c$ opens the backdoor path $z \leftarrow c \rightarrow y$, and the perturbation is designed to break that path. The perturbation network is meta-learned on a small clean, balanced meta dataset augmented with counterfactual images (foreground removed, label flipped to non-$y$) and factual images (background removed, label kept), which encode the human knowledge that backgrounds should not determine labels. The paper reports consistent state-of-the-art results on long-tail, noisy-label, subpopulation-shift, and generalized long-tail benchmarks, and its GradCAM visualizations show model attention moving from background to foreground.

Load-bearing premise

The method assumes the mask that separates object from background is correct: the counterfactual images must genuinely remove the object and the factual images must genuinely keep it, otherwise the meta-learning signal teaches the perturbation network to rely on wrong causal information.

Editorial extensions

If this is right

  • If the intervention story is correct, CLP should improve worst-group accuracy under subpopulation shift, not just average accuracy; the paper reports this on Waterbirds, with worst-group accuracy rising from 90.6% to 91.8%.
  • Because the perturbation network consumes training characteristics rather than hand-coded class-level rules, one training procedure covers long-tail, noisy-label, subpopulation-shift, and generalized long-tail settings; the paper reports top results in all four.
  • The causal augmentation is the active ingredient: the paper's Meta-LP variant, identical but without counterfactual and factual augmentation, performs worse on every benchmark, indicating that the gains come from the injected causal knowledge and not from the meta-learning loop alone.
  • Only the meta dataset, not the full training set, is causally augmented, so the extra compute over prior meta-learning methods is modest; the paper's per-epoch timing comparison shows CLP is comparable to Meta-Weight-Net and MetaSAug.

Reading between the lines

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

  • Editorial inference: deliberately corrupting foreground masks would likely erode CLP's gains, making mask quality the practical bottleneck; the paper does not report this ablation.
  • Editorial inference: the same meta-learning recipe should transfer to other non-causal attributes such as color, texture, or a protected attribute in tabular data, whenever a mask-like annotation can be produced; the paper only sketches the color case.
  • Editorial inference: because Meta-LP already beats many baselines, part of CLP's edge comes from the clean balanced meta set itself, so the causal metadata is an additive layer rather than the sole source of robustness.
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 / 4 minor

Summary. The paper introduces Causal Logit Perturbation (CLP), a meta-learning framework that trains a classifier with sample-wise logit perturbations. A two-layer MLP generates additive logit perturbations from ten training characteristics of each sample, and is optimized on a small clean metadata set augmented in two ways: counterfactual images (foreground infilled, labeled 'non-y') and factual images (background infilled, label preserved). The authors claim that this procedure implements an intervention do(u_z) on background-driven logit components, thereby closing the backdoor path z ← c → y. The method is evaluated on long-tail CIFAR, noisy CIFAR, Waterbirds, and generalized long-tail ImageNet/MSCOCO benchmarks, with an ablation called Meta-LP that removes the causal augmentation. The empirical tables show consistent improvements over the reported baselines, and the authors include ablation studies for the saliency term, augmentation styles, lambda, metadata size, and the ten characteristics.

Significance. If the causal-intervention interpretation were established, CLP would be a broadly applicable recipe: a small clean metadata set plus foreground masks plus a learned logit perturbation would mitigate background-label spurious correlations across many benchmark settings. The paper has genuine strengths: the Meta-LP ablation isolates the causal augmentation, the ablation suite is fairly thorough, and the method is computationally comparable to prior meta-learning baselines. However, the central causal claim, as formalized in Eq. (1), is not actually enforced or verified by the design; the empirical gains could plausibly come from standard meta-learned logit adjustment driven by class proportion, sample difficulty, and clean balanced metadata. The paper also reports no variance over its three seeds, and it never validates the foreground masks on which the counterfactual labels and saliency regularization depend. These gaps prevent the causal interpretation from being accepted as stated, although the empirical method itself may be defensible after reframing.

major comments (3)
  1. [Section 3.1, Eq. (1)] Equation (1) states that the perturbation p(u) = f_U(do(u_z), u_y), but the actual perturbation is an unconstrained additive vector δ_x = MLP(g_x; Ω) added to the full logit vector. Nothing in the bi-level objective (Eq. (7)) or the updates (Eqs. (8)-(10)) constrains δ to act only on the extrinsic, background-driven component u_z or to leave u_y invariant. The meta-objective only requires low loss on the augmented metadata, including the non-specific counterfactual loss in Eq. (3). Because the perturbation network receives class proportion (g6), average loss (g7), and relative margin (g10) as inputs, the observed gains can be reproduced by standard class-imbalance or difficulty logit adjustments without closing the z ← c → y backdoor path. The causal interpretation therefore does not follow from the design; the authors should either add a mechanism or test that verifies the u_z-only effect, or restate the contribution as a meta-learned logit adjustment guided by causal-style metadata augmentation.
  2. [Section 3.2.1, Eqs. (2)-(4)] The counterfactual supervision is under-specified. Counterfactually augmented images are labeled 'non-y' and trained with ℓ_CF = -log(1 - P(ŷ = y | x)) (Eq. (3)), which only forces the probability of the original class toward zero; it does not specify a positive alternative class. This can be satisfied by suppressing confidence or by increasing any other class, making the gradient signal weak and model-dependent. In addition, the correctness of the 'non-y' label depends on the HAttMatting masks and bounding boxes described in Section 3.2.1, but mask quality and the semantic validity of the infilled counterfactuals are never evaluated. If a mask fails to cover the foreground, the augmented label is wrong, and the saliency regularizer in Eq. (6) penalizes the wrong pixels. The paper should report mask-quality statistics (e.g., IoU against available annotations) and define a concrete positive-class objective for counterfactual samples.
  3. [Section 4, Tables 2, 5, 6, 7, 8] All headline results are reported as single numbers even though the experiments are run with three seeds. Many of the claimed improvements over the strongest baselines are small, for example 1.43% on CIFAR10-LT 100:1 in Table 2, 0.98% on noisy CIFAR10 uniform in Table 5, and roughly 1-2% on the GLT protocols in Tables 7 and 8. Without standard deviations, per-seed values, or a significance test, the 'consistently achieves state-of-the-art' claim is not supported beyond seed noise. Reporting error bars, or at least per-seed results, is necessary to evaluate whether the gains are reliable.
minor comments (4)
  1. [Section 3.2.3, Eqs. (7)-(10)] The saliency term RSal in Eq. (6) requires a causal mask r for every image, and Eq. (8) applies it to training minibatches; however, the paper only explains mask generation for the metadata, not for the full training set. Please clarify whether HAttMatting masks are computed for all training samples, and if so, how their computational cost and errors are handled.
  2. [Section 3.2.2, Eq. (5)] The derivation of the cosine feature in Eq. (5) is abbreviated: the logit is equal to cos θ only after both weight and feature normalization, and a bias term is omitted. A short clarifying sentence would prevent confusion.
  3. [Section 4.1, Table 3] The augmentation-style comparison reports only the final error rate for each style, but the paper does not state which of the five counterfactual or four factual methods performed best on which imbalance ratio; a short recommendation, or a table split by imbalance ratio, would increase practical usefulness.
  4. [General] No code or implementation details (e.g., exact MLP hidden size, optimizer for the perturbation network, mask generation pipeline for CIFAR) are provided. Releasing code would substantially improve reproducibility, especially given the number of components in the framework.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity is found: CLP's gains rest on external benchmarks and a controlled ablation, not on a self-referential reduction.

full rationale

The derivation chain is self-contained. CLP's perturbation network is trained with a bi-level meta-objective on a clean metadata set augmented with counterfactual and factual samples (Eqs. 7-10), and the resulting classifier is evaluated on held-out test sets of four separate benchmarks, so the reported accuracy gains are not definitionally forced by the training objective. The controlled ablation Meta-LP, which removes exactly the causal augmentation, isolates the contribution of the causal metadata rather than folding it into the method's definition. Eq. (1) is a stated design target for what a causal logit perturbation should do, not a quantity fitted to the test results, and the paper does not claim to predict test accuracy from Eq. (1). The saliency regularization (Eq. 6) does encode the same causal prior as the evaluation (backgrounds should be ignored), but it is a component of the training loss and the GradCAM visualizations are a direct check of that regularizer, not a prediction derived from it. Self-citations (e.g., refs. 16, 26-29) are to prior logit-adjustment and meta-learning work and are not load-bearing; no uniqueness theorem or externally cited first-principles result is replaced by a same-author claim. The skeptic's concern that the unconstrained additive perturbation may not literally implement do(u_z) is an interpretive gap about the mechanism, not a circular reduction: nothing in the paper defines the perturbation as the intervention and then re-derives the intervention from it. Hence no circular step meets the evidentiary bar.

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

The method rests on a small clean meta set, on the assumption that foregrounds are causal and backgrounds are spurious, and on the assumption that a learned additive logit perturbation can stand in for a do-calculus intervention. The counterfactual 'non-y' target is an ad hoc label. No new physical or conceptual entities are introduced.

free parameters (4)
  • Saliency regularization strength lambda = selected from {0.2, 0.4, 0.6, 0.8, 1}
    Controls how strongly the classifier is penalized for attending to background pixels; tuned per dataset and discussed in Section 4.6.
  • Augmented metadata size and CF:F ratio = 2x metadata size, 1:1 counterfactual:factual
    Number of augmented samples and the ratio are set by hand; the sensitivity analysis in Section 4.6 shows performance stabilizes above a 10:1 augmentation ratio.
  • Number of KMeans clusters for metadata construction = 6
    Used to build attribute-balanced metadata for ImageNet-GLT and MSCOCO-GLT in Section 4.4.
  • FGSM perturbation epsilon = 0.5
    Factual augmentation uses FGSM with the infinity norm and epsilon 0.5 to perturb backgrounds, chosen by hand in Section 3.2.1.
assumptions (5)
  • domain assumption Additive logit decomposition and the claim that logit perturbation can realize do(uz)
    Eq. (1) states the perturbed logit is f_U(do(uz), uy), but no derivation shows that adding a learned delta is equivalent to a do-calculus intervention on the extrinsic logit component.
  • domain assumption Foreground is the causal attribute and background is non-causal
    The entire augmentation pipeline and the saliency term in Eq. (6) rely on masks and bounding boxes from HAttMatting; no validation of mask quality is provided.
  • domain assumption Metadata is clean, balanced, and representative
    The perturbation network is meta-trained on a small trusted set drawn from validation or training data; if labels in the metadata are not clean, the learned perturbations inherit the bias.
  • ad hoc to paper The ten training characteristics are sufficient to generate useful perturbations
    Table 1 lists ten features borrowed from prior work; no theoretical argument is given that this feature set spans the relevant training dynamics.
  • ad hoc to paper Counterfactual label 'non-y' is a valid training target
    Eq. (3) replaces the original label with the set of all other classes; a background-only image is not necessarily 'not y', so this is a weak and potentially incorrect supervision signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Spurious Correlations with Causal Logit Perturbation." pith.science (2026). https://pith.science/paper/EIQWE4WJ

@misc{pith2026250515246,
  author       = {Pith},
  title        = {Pith review of: Mitigating Spurious Correlations with Causal Logit Perturbation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EIQWE4WJ}},
  note         = {Machine review of arXiv:2505.15246}
}
read the original abstract

Deep learning has seen widespread success in various domains such as science, industry, and society. However, it is acknowledged that certain approaches suffer from non-robustness, relying on spurious correlations for predictions. Addressing these limitations is of paramount importance, necessitating the development of methods that can disentangle spurious correlations. {This study attempts to implement causal models via logit perturbations and introduces a novel Causal Logit Perturbation (CLP) framework to train classifiers with generated causal logit perturbations for individual samples, thereby mitigating the spurious associations between non-causal attributes (i.e., image backgrounds) and classes.} {Our framework employs a} perturbation network to generate sample-wise logit perturbations using a series of training characteristics of samples as inputs. The whole framework is optimized by an online meta-learning-based learning algorithm and leverages human causal knowledge by augmenting metadata in both counterfactual and factual manners. Empirical evaluations on four typical biased learning scenarios, including long-tail learning, noisy label learning, generalized long-tail learning, and subpopulation shift learning, demonstrate that CLP consistently achieves state-of-the-art performance. Moreover, visualization results support the effectiveness of the generated causal perturbations in redirecting model attention towards causal image attributes and dismantling spurious associations.

Figures

Figures reproduced from arXiv: 2505.15246 by the authors.

Figure 1
Figure 1. Illustration for spurious correlation resulting from a rare background occurrence. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The causal graph before (a) and after (b) the intervention process of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The overall structure of CLP, which consists of four main components: the meta [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Class-wise accuracy on CIFAR10-LT with imbalance ratios of 10:1 (a) and 100:1 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Ratio of samples in CIFAR10 where perturbations from noise and clean samples [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the regions that the model used for making predictions. The [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Ablation studies for the saliency regularization term on CIFAR10 (a) and CI [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Sensitivity tests of λ on CIFAR10 (a) and CIFRA100 (b) datasets with 20% flip noise. Sensitivity test of the number of augmented samples on CIFAR10 (c) and CIFAR100 (d) datasets with 40% flip label noise. The horizontal axis represents the ratio of the number of augmen…
Figure 9
Figure 9. Figure 9: Ablation studies of extracted training characteristics under LT learning scenarios [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 51 canonical work pages

  1. [1]

    Invariant risk minimization

    Arjovsky M, Bottou L, Gulrajani I, et al. Invariant risk minimization. 2019, arXiv: 1907.02893. 28

    arXiv 2019

  2. [2]

    Invariance principle meets infor- mation bottleneck for out-of-distribution generalization

    Ahuja K, Caballero E, Zhang D, et al. Invariance principle meets infor- mation bottleneck for out-of-distribution generalization. In: Proceed- ings of the 35th Advances in Neural Information Processing Systems (NeurIPS), 2021. 3438-3450

    work page 2021

  3. [3]

    Out-of-distribution gener- alization via risk extrapolation (REx)

    Krueger D, Caballero E, Jacobsen J H, et al. Out-of-distribution gener- alization via risk extrapolation (REx). In: Proceedings of the 38th In- ternational Conference on Machine Learning (ICML), 2021. 5815-5826

    work page 2021

  4. [4]

    Deep CORAL: Correlation alignment for deep domain adaptation

    Sun B, Saenko K. Deep CORAL: Correlation alignment for deep domain adaptation. In: Proceedings of the European Conference on Computer Vision (ECCV), 2016. 443-450

    work page 2016

  5. [5]

    Counterfactual generative networks

    Sauer A, Geiger A. Counterfactual generative networks. In: Proceed- ings of the 9th International Conference on Learning Representations (ICLR), 2021

    work page 2021

  6. [6]

    Towards robust classification model by counterfactual and invariant data generation

    Chang C H, Adam G A, and Anna G. Towards robust classification model by counterfactual and invariant data generation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 15212-15221

    work page 2021

  7. [7]

    Long-tail learning via logit adjustment

    Menon A K, Jayasumana S, Rawat A S, et al. Long-tail learning via logit adjustment. In: Proceedings of the 8th International Conference on Learning Representations (ICLR), 2020

    work page 2020

  8. [8]

    Learning imbalanced datasets with label- distribution-aware margin loss

    Cao K, Wei C, Gaidon A, et al. Learning imbalanced datasets with label- distribution-aware margin loss. In: Proceedings of the 32nd Advances in Neural Information Processing Systems (NeurIPS), 2019. 1567–1578

    work page 2019

Show all 54 references
  1. [9]

    Class-level logit perturbation

    Li M, Su F, Wu O, et al. Class-level logit perturbation. IEEE Trans. Neural. Networks Learn. Sys., 2023, 1-15

    2023

  2. [10]

    Causality for machine learning

    Scholkopf B. Causality for machine learning. 2021, arXiv: 1911.10500

    2021 arXiv

  3. [11]

    Correct-N-Contrast: a con- trastive approach for improving robustness to spurious correlations

    Zhang M, Sohoni N S, Zhang H R, et al. Correct-N-Contrast: a con- trastive approach for improving robustness to spurious correlations. In: Proceedings of the 39th International Conference on Machine Learning (ICML), 2022. 26484-26516. 29

    2022

  4. [12]

    Augmentation by counterfac- tual explanation-fixing an overconfident classifier

    Singla S, Murali N, Arabshahi F, et al. Augmentation by counterfac- tual explanation-fixing an overconfident classifier. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV), 2023. 4720-4730

    2023

  5. [13]

    Robust learning with pro- gressive data expansion against spurious correlation

    Deng Y, Yang Y, Mirzasoleiman B, et al. Robust learning with pro- gressive data expansion against spurious correlation. In: Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS), 2023. 1390-1402

    2023

  6. [14]

    Generative interventions for causal learning

    Mao C, Cha A, Gupta A, et al. Generative interventions for causal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3947–3956

    2021

  7. [15]

    Adjusting logit in Gaussian form for fong-tailed visual recognition

    Li M, Cheung Y, Lu Y, et al. Adjusting logit in Gaussian form for fong-tailed visual recognition. IEEE. Trans. Artif. Intell., 2024, 5(10): 5026-5039

    2024

  8. [16]

    Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail Learning

    Zhou X, Wu O, Yang N. Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail Learning. In: Proceedings of the AAAI Con- ference on Artificial Intelligence (AAAI). 2025. 22991-22999

    2025

  9. [17]

    Adaptive logit adjustment loss for long- tailed visual recognition

    Zhao Y, Chen W, Tan X, et al. Adaptive logit adjustment loss for long- tailed visual recognition. In: Proceedings of the 36th AAAI Conference on Artificial Intelligence (AAAI), 2022. 3472-3480

    2022

  10. [18]

    Imagine by reasoning: A reasoning- based implicit semantic data augmentation for long-tailed classification

    Chen X, Zhou Y, Wu D, et al. Imagine by reasoning: A reasoning- based implicit semantic data augmentation for long-tailed classification. In: Proceedings of the 36th AAAI Conference on Artificial Intelligence (AAAI), 2022. 356-364

    2022

  11. [19]

    Model-agnostic meta-learning for fast adap- tation of deep networks

    Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adap- tation of deep networks. In: Proceedings of the 34th International Con- ference on Machine Learning (ICML), 2017. 1126-1135

    2017

  12. [20]

    MetaSAug: Meta semantic augmentation for long-tailed visual recognition

    Li S, Gong K, Liu C H, et al. MetaSAug: Meta semantic augmentation for long-tailed visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5212-5221. 30

    2021

  13. [21]

    Meta-Weight-Net: Learning an explicit map- ping for sample weighting

    Shu J, Xie Q, Yi L, et al. Meta-Weight-Net: Learning an explicit map- ping for sample weighting. In: Proceedings of the 33rd Advances in Neural Information Processing Systems (NeurIPS), 2019. 1919–1930

    2019

  14. [22]

    Focal loss for dense object de- tection

    Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object de- tection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 2980-2988

    2017

  15. [23]

    Long-tailed recognition via weight balancing

    Alshammari S, Wang Y X, Ramanan D, et al. Long-tailed recognition via weight balancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6897-6907

    2022

  16. [24]

    Attention-guided hierarchical structure aggregation for image matting

    Qiao Y, Liu Y, Yang X, et al. Attention-guided hierarchical structure aggregation for image matting. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), 2020. 13676-13685

    2020

  17. [25]

    Class-balanced loss based on effective number of samples

    Cui Y, Jia M, Lin T Y, et al. Class-balanced loss based on effective number of samples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 9268-9277

    2019

  18. [26]

    Combining adversaries with anti-adversaries in training

    Zhou X, Yang N, Wu O. Combining adversaries with anti-adversaries in training. In: Proceedings of the AAAI Conference on Artificial Intelli- gence (AAAI), 2023. 11435-11442

    2023

  19. [27]

    Implicit counterfactual data augmentation for deep neu- ral networks

    Zhou X, Wu O. Implicit counterfactual data augmentation for deep neu- ral networks. arXiv e-prints, 2023: arXiv: 2304.13431

    2023 arXiv

  20. [28]

    Valuing training data via causal inference for in-context learning

    Zhou X, Ye W, Lee Z, et al. Valuing training data via causal inference for in-context learning. IEEE Trans. Knowl. Data Eng., 2025

    2025

  21. [29]

    Boosting model resilience via implicit ad- versarial data augmentation

    Zhou X, Ye W, Lee Z, et al. Boosting model resilience via implicit ad- versarial data augmentation. In: Proceedings of the Thirty-Third In- ternational Joint Conference on Artificial Intelligence (IJCAI). 2024: 5653-5661

    2024

  22. [30]

    Large-scale long-tailed recognition in an open world

    Liu Z, Miao Z, Zhan X, et al. Large-scale long-tailed recognition in an open world. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2537-2546. 31

    2016

  23. [31]

    Deep residual learning for image recogni- tion

    He K, Zhang X, Ren S, et al. Deep residual learning for image recogni- tion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 770-778

    2016

  24. [32]

    Implicit semantic data augmentation for deep networks

    Wang Y, Pan X, Song S, et al. Implicit semantic data augmentation for deep networks. In: Proceedings of the 32nd Advances in Neural Information Processing Systems (NeurIPS), 2019. 12635–12644

    2019

  25. [33]

    How re-sampling helps for long- tail learning? In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS)

    Shi, J X, Wei, T, Xiang, Y, et al. How re-sampling helps for long- tail learning? In: Proceedings of the Advances in Neural Information Processing Systems (NeurIPS). 2023, 75669–75687

    2023

  26. [34]

    Learning multiple layers of features from tiny images

    Krizhevsky A, Hinton G. Learning multiple layers of features from tiny images. Technical Reports. 2009

    2009

  27. [35]

    Wide residual networks

    Zagoruyko S, Komodakis N. Wide residual networks. In: Proceedings of the British Machine Vision Conference 2016 (BMVC), 2016

    2016

  28. [36]

    Co-teaching: Robust training of deep neu- ral networks with extremely noisy labels

    Han B, Yao Q, Yu X, et al. Co-teaching: Robust training of deep neu- ral networks with extremely noisy labels. In: Proceedings of the 32nd Advances in Neural Information Processing Systems (NeurIPS), 2018. 8536-8546

    2018

  29. [37]

    Normalized loss functions for deep learning with noisy labels

    Ma X, Huang H, Wang Y, et al. Normalized loss functions for deep learning with noisy labels. In: Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 6543-6553

    2020

  30. [38]

    Combating noisy labels by agreement: A joint training method with co-regularization

    Wei H, Feng L, Chen X, et al. Combating noisy labels by agreement: A joint training method with co-regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 13726-13735

    2020

  31. [39]

    L DMI: A novel information-theoretic loss function for training deep nets robust to label noise

    Xu Y, Cao P, Kong Y, et al. L DMI: A novel information-theoretic loss function for training deep nets robust to label noise. In: Proceedings of the 33rd Advances in Neural Information Processing Systems (NeurIPS),

  32. [40]

    MentorNet: Learning data-driven cur- riculum for very deep neural networks on corrupted labels

    Jiang L, Zhou Z, Leung T, et al. MentorNet: Learning data-driven cur- riculum for very deep neural networks on corrupted labels. In: Proceed- ings of the 35th International Conference on Machine Learning (ICML),

  33. [41]

    Learning to reweight examples for robust deep learning

    Ren M, Zeng W, Yang B, et al. Learning to reweight examples for robust deep learning. In: Proceedings of the 34th International Conference on Machine Learning (ICML), 2018. 4334-4343

    2018

  34. [42]

    Using trusted data to train deep networks on labels corrupted by severe noise

    Hendrycks D, Mazeika M, Wilson D, et al. Using trusted data to train deep networks on labels corrupted by severe noise. In: Proceedings of the 32nd Advances in Neural Information Processing Systems (NeurIPS),

  35. [43]

    Meta-learning advisor networks for long-tail and noisy labels in social image classification

    Ricci S, Uricchio T, and Bimbo A. Meta-learning advisor networks for long-tail and noisy labels in social image classification. ACM Trans. Multimedia Comput. Commun. Appl., 2023, 19(5s): 169

    2023

  36. [44]

    Distributionally robust neural networks

    Sagawa S, Koh P W, Hashimoto T B, et al. Distributionally robust neural networks. In Proceedings of the 7th International Conference on Learning Representations (ICLR), 2019

    2019

  37. [45]

    The Caltech-UCSD Birds-200-2011 dataset

    Wah C, Branson S, Welinder P, et al. The Caltech-UCSD Birds-200-2011 dataset. California Institute of Technology, 2021

    2011

  38. [46]

    Places: A 10 million image database for scene recognition

    Zhou B, Lapedriza A, Khosla A, et al. Places: A 10 million image database for scene recognition. IEEE Trans. Pattern Anal. Mach. In- tell., 2017, 40: 1452-1464

    2017

  39. [47]

    Improving out-of-distribution robustness via selective augmentation

    Yao H, Wang Y, Li S, et al. Improving out-of-distribution robustness via selective augmentation. In: Proceedings of the 39th International Conference on Machine Learning (ICML), 2022. 25407-25437

    2022

  40. [48]

    Adversarial domain adaptation with do- main mixup

    Xu M, Zhang J, Ni B, et al. Adversarial domain adaptation with do- main mixup. In: Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI), 2020. 6502-6509

    2020

  41. [49]

    Invariant feature learning for generalized long-tailed classification

    Tang K, Tao M, Qi J, et al. Invariant feature learning for generalized long-tailed classification. In: Proceedings of the European Conference on Computer Vision (ECCV), 2022. 709-726

    2022

  42. [50]

    Aggregated residual transformations for deep neural networks

    Xie S, Girshick R, Doll´ ar P, et al. Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2017. 1492- 1500.4 33

    2017

  43. [51]

    Decoupling representation and classi- fier for long-tailed recognition

    Kang B, Xie S, Rohrbach M, et al. Decoupling representation and classi- fier for long-tailed recognition. In: Proceedings of the 7th International Conference on Learning Representations (ICLR), 2019

    2019

  44. [52]

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

    Tang K, Huang J, Zhang H. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In: Proceedings of the 34th Advances in Neural Information Processing Systems (NeurIPS),

  45. [53]

    Balanced meta-softmax for long-tailed visual recognition

    Ren J, Yu C, Ma X, et al. Balanced meta-softmax for long-tailed visual recognition. In: Proceedings of the 34th Advances in Neural Information Processing Systems (NeurIPS), 2020. 4175-4186

    2020

  46. [54]

    RandAugment: Practical automated data augmentation with a reduced search space

    Cubuk E D, Zoph B, Shlens J, et al. RandAugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3008-3017. 34

    2020

Pith tools

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