Pith. sign in

REVIEW 4 major objections 3 minor 60 references

NeuronTune: Towards Self-Guided Spurious Bias Mitigation

T0 review · 4 major / 3 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 NeuronTune claims that the median activation gap between a model's correct and incorrect predictions exposes the neurons carrying spurious correlations, and that zeroing these neurons during last-layer retraining moves the model closer to…

desk verdict Genuinely useful empirical recipe for post hoc debiasing, but the advertised theoretical guarantee rests on a proof that does not hold up; worth peer review only with a major revision. read the letter →

arxiv 2505.24048 v1 pith:VZM5Y6H5 submitted 2025-05-29 cs.LG

Guangtao Zheng , Wenqian Ye , Aidong Zhang This is my paper
classification cs.LG
keywords spuriouscorrelationworst-groupaccuracyneuronselectionlast-layerretrainingposthocdebiasingself-guidedbiasmitigationlatentembeddinganalysis
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

NeuronTune claims that a trained model's own prediction errors point to the neurons that carry spurious correlations. For each latent neuron and class, it compares the median activation on misclassified examples against the median on correctly classified ones; when misclassified examples activate a neuron more strongly, the neuron is flagged as biased. The paper proves, in a linear data model, that this median gap tracks the neuron's weight on the spurious feature, and that zeroing the flagged neurons while retraining only the last layer moves the model closer to the unbiased solution. If correct, this gives a cheap, post-hoc way to raise worst-group accuracy without any group labels.

What carries the argument

The load-bearing object is the spuriousness score $\delta_i^y = \mathrm{Med}(\text{incorrect activations}) - \mathrm{Med}(\text{correct activations})$ computed per latent neuron per class, with the selection set $S = \{i \mid \delta_i^y > \lambda\}$. In the paper's linear model $x = x_{\mathrm{core}} \oplus x_{\mathrm{spu}}$ with $x_{\mathrm{spu}} = (2a-1)\gamma y + \varepsilon_{\mathrm{spu}}$, the identity $\delta_i^y \approx -2\mu\,\gamma^\top w_{\mathrm{spu},i}$ makes the median gap a readable proxy for how strongly a neuron encodes the spurious direction. NeuronTune feeds the identified set into a class-balanced last-layer retraining objective, masking those dimensions to zero, and selects the tuning epoch through the spuriousness fitness score $\mathrm{SFit} = \sum_m \sum_y |\delta_m^y|$.

What would settle it

On the Waterbirds validation split (467 landbird-on-land versus 466 landbird-on-water), compute the fraction of misclassified landbird samples that come from the minority group; if that fraction is not above one half, Lemma 2's premise fails, and the decisive test would be whether suppressing randomly chosen neurons by last-layer retraining matches NeuronTune's worst-group accuracy.

Watch

Extended reading notes

Core claim

The central claim is that spurious bias can be located inside a trained model by a simple statistic: for each latent dimension $i$ and class $y$, $\delta_i^y = \mathrm{Med}(\bar{V}_i^y) - \mathrm{Med}(\hat{V}_i^y)$, the median activation of misclassified samples minus the median activation of correctly predicted samples. The paper proves (Theorem 4.2) that under its linear data model $\delta_i^y \approx -2\mu\,\gamma^\top w_{\mathrm{spu},i}$, so a positive $\delta_i^y$ selects exactly the neurons whose embedding weights oppose the spurious feature direction. Retraining the final linear layer with those neurons zeroed out leaves the weight on the spurious component unchanged (Lemma 3) but pulls the core weight closer to the optimal $\beta$ (Theorem 4.3), producing a model provably closer to the unbiased one than the ERM baseline. Empirically, NeuronTune improves worst-group accuracy on Waterbirds, CelebA, MultiNLI, CivilComments, and ImageNet-A without group annotations.

Load-bearing premise

The guarantee that misclassified samples come mostly from the minority spurious group assumes the majority group covers more than three-quarters of the identification data with high-confidence predictions, a condition that does not hold on the balanced validation sets used in the experiments.

Editorial extensions

If this is right

  • Applying NeuronTune to an ERM-trained model improves worst-group accuracy on Waterbirds, CelebA, MultiNLI, and CivilComments without group labels, giving the best unsupervised results reported in Tables 1 and 2.
  • NeuronTune reduces the gap between average accuracy and worst-group accuracy, for example on Waterbirds from 24.7 to 2.2, indicating lower reliance on spurious correlations.
  • On the ImageNet-9 to ImageNet-A transfer, NeuronTune reaches the best ImageNet-A accuracy (37.3) among compared methods, showing robustness to distribution shifts that group labels cannot easily describe.
  • The spuriousness fitness score SFit acts as a model-selection metric in the unsupervised setting, so hyperparameter choice needs no group labels.
  • Full suppression of the identified neurons (masking value 0) is necessary; partial suppression fails to improve worst-group accuracy, as shown in Table 5.

Reading between the lines

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

  • A natural testable extension is to replace the medians with quantile- or variance-weighted statistics on high-dimensional embeddings, whose richer geometry may give a more reliable neuron-selection signal when the balanced-data premise fails.
  • Because the method assumes a frozen feature extractor and a linear last layer, it can be applied directly to foundation-model embeddings, where its linear-probe assumptions match the practical setup.
  • The same $\delta$ statistic could serve as a pre-mitigation diagnostic: a high spuriousness score would flag that a model's high average accuracy hides shortcut reliance, even before any debiasing is applied.
  • If the median-difference signal is a general signature of shortcut learning, combining NeuronTune with iterative identification across multiple layers could address bias that is entangled deeper than the final embedding.
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 / 3 minor

Summary. The paper proposes NeuronTune, a post hoc method for spurious-bias mitigation that requires no group labels. The method estimates a spuriousness score for each embedding dimension as the difference between the median activation of misclassified samples and the median activation of correctly classified samples, suppresses the dimensions with positive scores, and retrains the last layer on class-balanced data. The authors present an analytical framework with a linear-regression data model, a neuron-selection principle (Proposition 4.1), a selection metric (Theorem 4.2), and a guarantee that the procedure moves the model closer to an unbiased classifier (Theorem 4.3). Experiments on Waterbirds, CelebA, ImageNet-9/A, MultiNLI, and CivilComments show consistent worst-group-accuracy improvements over several unsupervised and semi-supervised baselines, with ablations on the identification data and masking strategy.

Significance. If the theoretical claims were sound, the paper would make a useful contribution: neuron-level, self-guided intervention is a plausible and efficient alternative to sample-level reweighting, and the reported WGA gains across vision and text datasets, with code provided, suggest the heuristic has practical value. However, the central theoretical guarantee in Theorem 4.3 is not established: the proof contains an invalid norm/projection step and an unjustified equality that carry the entire 'closer to unbiased' conclusion. In addition, the practical configuration (Dval as identification data, absolute activations) deviates from the assumptions used in the theory, and the most closely related unsupervised baseline, EVA (He et al., 2025), is cited but not compared. The empirical contribution may still stand as a heuristic, but the advertised theoretical backing and the comparative evaluation require substantial revision.

major comments (4)
  1. [Appendix A.2.6 / Section 4.1.4] The proof of Theorem 4.3 has two load-bearing steps that do not follow. First, Eq. (32) asserts γ^T u†_spu = z* after re-optimizing b only over the retained set I+; changing the feasible set by zeroing neurons alters the optimality condition, and the equality with the unconstrained optimum z* is not derived. The subsequent claim that b†_i > 0 for all i in I+ also does not follow from a positive weighted sum. Second, Eq. (33) equates ∥u†_core − β∥² with ∥β^T(u†_core − β)∥²; this holds only if u†_core − β is parallel to β, which is not shown. Consequently the chain ending at ∥z* − 1∥ bounds only one component of the distance and does not establish that u†_core is closer to β than the biased solution (1−z*)β. Because Theorem 4.3 is the sole source of the 'guaranteed to outperform ERM' and 'closer to unbiased' statements, the central theoretical claim is unsupported as written.
  2. [Appendix A.2.7 / Section 5.2] Lemma 2, which underpins Theorem 4.2, requires p > 3/4 (and a high signal-to-noise ratio) so that a majority of correctly predicted samples come from a=1 and a majority of incorrectly predicted samples come from a=0. The practical method uses the validation set as DIde, and in the reported experiments those validation groups are roughly balanced within classes (e.g., Waterbirds val has 467 vs 466 landbird samples and 133 vs 133 waterbird samples; CelebA non-blond val has 8535 vs 8276). With p ≈ 0.5, Lemma 2's conclusion fails, so the median-difference signal δ_i^y is not guaranteed to align with γ^T w_spu,i < 0. The theory therefore does not justify the selection metric in the actual configuration where the method performs best.
  3. [Section 5.2] The implementation computes the spuriousness score on the absolute values of neuron activations ('We took the absolute values of neuron activations before the identification process'), whereas Theorem 4.2 defines activation values as x_core^T w_core,i + x_spu^T w_spu,i and proves δ_i^y ≈ −2µ γ^T w_spu,i for these signed activations. The proof does not carry over to absolute activations, and the sign structure that links high incorrect activations to negative γ^T w_spu,i can be altered by taking absolute values. This is a direct mismatch between the theoretical justification and the evaluated method.
  4. [Section 2 / Tables 1-3] The closest prior work to NeuronTune is EVA (He et al., 2025), which is cited in Section 2 as an unsupervised method that mitigates spurious bias by erasing activations with high confidence. EVA is not included in any of the comparison tables. Because EVA operates in the same setting and uses a similar activation-erasure mechanism, the absence of this baseline makes it difficult to assess the marginal benefit of the median-difference selection criterion, which is the main novel component of NeuronTune.
minor comments (3)
  1. [Appendix A.2.6] The final display line of the proof drops the square and the norm notation: it should state ∥u†_core − β∥ ≈ |z* − 1|, not ∥z* − 1∥, after the square is taken.
  2. [Section 5.4] The sentence 'NeuronTune in theory can achieve better robustness to spurious bias (Appendix A.3)' overstates the current theoretical support, since Theorem 4.3's proof is invalid as written; the sentence should be revised to reflect the empirical comparison only.
  3. [Section 4.2.1] The criterion uses a single threshold λ across all classes y, while the metric δ_i^y is defined per class; the paper should clarify whether the threshold is applied to each class separately and whether the union set S across classes is intended to be interpreted dimension-wise despite class-specific scores.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the neuron-selection metric and the closeness guarantee are derived from the stated generative model, not fitted to the target; self-citations are contextual.

full rationale

The derivation chain is self-contained with respect to circularity. Lemma 1 solves the ERM objective under the stated data model (x_spu = (2a-1)gamma y + epsilon_spu) to obtain u*_core and u*_spu; Theorem 4.2 derives delta_i^y approximately -2 mu gamma^T w_spu,i from that model plus Lemma 2, and Proposition 4.1 defines spurious neurons as those with gamma^T w_spu,i < 0, so the selection criterion follows from the model rather than being fitted to worst-group accuracy. Theorem 4.3 is a conditional statement: under the stated assumption beta^T w_core,i approximately gamma^T w_spu,i, it argues that NeuronTune moves u_core toward beta. The assumption is explicitly stated, not defined in terms of the conclusion, and no parameter is fitted to the quantity being predicted. The appendix proof contains unjustified equalities, e.g., Eq. 33 equating a full L2 norm with a projection onto beta, and Eq. 32 asserting u_dagger_spu = u*_spu without a complete derivation; these are proof-validity concerns rather than circularity in the sense of a reduction to the input or a self-citation chain. Self-citations (Ye et al. 2024; Zheng et al. 2024b,c; 2025) are contextual references and are not load-bearing for the theoretical claims. The experimental evaluation is against external benchmarks (Waterbirds, CelebA, ImageNet-9/A, MultiNLI, CivilComments).

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper does not invent new physical or model entities; its new concepts are the spuriousness score δ_i^y and the biased-dimension set S. The main external inputs are the linear data model, the linear prediction model, and the two load-bearing technical conditions (Lemma 2's imbalance condition and Theorem 4.3's alignment assumption). The threshold λ and masking value are hand-chosen hyperparameters.

free parameters (2)
  • Threshold λ for biased dimension selection = 0 (default)
    Eq. (6) selects dimensions with δ_i^y > λ; λ=0 is set by hand and the paper says it 'works well in practice', with no theory fixing the value.
  • Masking value for suppressing biased dimensions = 0
    Section 5.5 and Table 5 test masking values 0 to 1.0; full suppression (0) is chosen as default because partial masking does not improve WGA on CelebA.
assumptions (5)
  • domain assumption Data model: x = x_core ⊕ x_spu, x_spu = (2a-1)γy + ε, with η_core^2 ≫ η_spu^2
    Eqs. (1)-(2). Standard in prior work (Arjovsky et al. 2019; Ye et al. 2023); assumed throughout the theory, requiring spurious features to be linear, separable, and predominantly correlated with the label.
  • domain assumption Prediction model f(x) = b^T W x (two linear layers)
    Eq. (3). The practical method uses deep nonlinear networks, so this axiom is a simplification that does not directly transfer to the experimental setting.
  • domain assumption Lemma 2 conditions: p>3/4 and high signal-to-noise ratio Φ(-μ/σ1)≈0
    Appendix A.2.7. Needed so that most misclassified samples are from the minority spurious group. The practical identification data Dval is roughly balanced, so this condition is questionable.
  • ad hoc to paper β^T w_core,i ≈ γ^T w_spu,i for all i
    Theorem 4.3 and Eqs. (35)-(36). This 'mild assumption' is load-bearing for concluding u†_core≈β; the paper asserts it 'generally holds' without evidence.
  • domain assumption Gaussian approximation for medians in Theorem 4.2
    Appendix A.2.5. Medians are replaced by expectations under Gaussian assumptions on activations; not verified on real embeddings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeuronTune: Towards Self-Guided Spurious Bias Mitigation." pith.science (2026). https://pith.science/paper/VZM5Y6H5

@misc{pith2026250524048,
  author       = {Pith},
  title        = {Pith review of: NeuronTune: Towards Self-Guided Spurious Bias Mitigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZM5Y6H5}},
  note         = {Machine review of arXiv:2505.24048}
}
read the original abstract

Deep neural networks often develop spurious bias, reliance on correlations between non-essential features and classes for predictions. For example, a model may identify objects based on frequently co-occurring backgrounds rather than intrinsic features, resulting in degraded performance on data lacking these correlations. Existing mitigation approaches typically depend on external annotations of spurious correlations, which may be difficult to obtain and are not relevant to the spurious bias in a model. In this paper, we take a step towards self-guided mitigation of spurious bias by proposing NeuronTune, a post hoc method that directly intervenes in a model's internal decision process. Our method probes in a model's latent embedding space to identify and regulate neurons that lead to spurious prediction behaviors. We theoretically justify our approach and show that it brings the model closer to an unbiased one. Unlike previous methods, NeuronTune operates without requiring spurious correlation annotations, making it a practical and effective tool for improving model robustness. Experiments across different architectures and data modalities demonstrate that our method significantly mitigates spurious bias in a self-guided way.

Figures

Figures reproduced from arXiv: 2505.24048 by the authors.

Figure 1
Figure 1. Practical implementation of NeuronTune. (a) Extract latent embeddings v1, . . . , vN and prediction outcomes (blue for correct and red for incorrect predictions) from an ERM-trained model using the identification data DIde. (b) Identify biased neurons (dimensions) utilizing the statistics Mmis and Mcor derived from neuron activations for correct (blue) and incorrect (red) predictions from Eq. (5). (c) Retrain the la… view at source ↗
Figure 2
Figure 2. Synthetic experiment. (a) Training and test data distri￾butions along with the decision boundaries of the trained model. (b) Value distributions of the correctly (blue) and incorrectly (red) predicted samples at the first (left) and second (right) dimensions of input embeddings, with the second dimension identified as a biased dimension. (c) NeuronTune improves WGA. Data groups (y = +1, a = 1): red dots; (y = +1, a … view at source ↗
Figure 3
Figure 3. Value distributions of the correctly (blue) and incorrectly (red) predicted samples for unbiased (a) and biased (b) dimensions, along with the representative samples, respectively, based on the non-blond hair samples in the CelebA dataset. (a) Identified unbiased dimensions for blond hair (b) Identified biased dimensions for blond hair [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Value distributions of the correctly (blue) and incorrectly (red) predicted samples for unbiased (a) and biased (b) dimensions, along with the representative samples, respectively, based on the blond hair samples in the CelebA dataset. 23 [PITH_FULL_IMAGE:figures/full…
Figure 5
Figure 5. Figure 5: Value distributions of the correctly (blue) and incorrectly (red) predicted samples for unbiased (a) and biased (b) dimensions, along with the representative samples, respectively, based on the landbird samples in the Waterbirds dataset. (a) Identified unbiased dimensi…
Figure 6
Figure 6. Figure 6: Value distributions of the correctly (blue) and incorrectly (red) predicted samples for unbiased (a) and biased (b) dimensions, along with the representative samples, respectively, based on the waterbird samples in the Waterbirds dataset. 24 [PITH_FULL_IMAGE:figures/f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Invariant risk minimization

    Arjovsky, M., Bottou, L., Gulrajani, I., and Lopez-Paz, D. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

    arXiv 1907

  3. [3]

    Bahng, H., Chun, S., Yun, S., Choo, J., and Oh, S. J. Learning de-biased representations with biased representations. In International Conference on Machine Learning, pp.\ 528--539. PMLR, 2020

    work page 2020

  4. [4]

    A., Dufumier, B., Tartaglione, E., Grangetto, M., and Gori, P

    Barbano, C. A., Dufumier, B., Tartaglione, E., Grangetto, M., and Gori, P. Unbiased supervised contrastive learning. In International Conference on Learning Representations, 2023

    work page 2023

  5. [5]

    Recognition in terra incognita

    Beery, S., Van Horn, G., and Perona, P. Recognition in terra incognita. In European Conference on Computer Vision, pp.\ 456--473, 2018

    work page 2018

  6. [6]

    Nuanced metrics for measuring unintended bias with real data for text classification

    Borkan, D., Dixon, L., Sorensen, J., Thain, N., and Vasserman, L. Nuanced metrics for measuring unintended bias with real data for text classification. In Companion proceedings of the 2019 world wide web conference, pp.\ 491--500, 2019

    work page 2019

  7. [7]

    R., and H \"o hne, M

    Bykov, K., Deb, M., Grinwald, D., Muller, K. R., and H \"o hne, M. M. Dora: Exploring outlier representations in deep neural networks. Transactions on Machine Learning Research, 2023

    work page 2023

  8. [8]

    and Lipton, Z

    Byrd, J. and Lipton, Z. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pp.\ 872--881. PMLR, 2019

    work page 2019

Show all 60 references
  1. [9]

    Rubi: Reducing unimodal biases for visual question answering

    Cadene, R., Dancette, C., Cord, M., Parikh, D., et al. Rubi: Reducing unimodal biases for visual question answering. In Advances in Neural Information Processing Systems, volume 32, 2019

    2019

  2. [10]

    Environment inference for invariant learning

    Creager, E., Jacobsen, J.-H., and Zemel, R. Environment inference for invariant learning. In International Conference on Machine Learning, pp.\ 2189--2200. PMLR, 2021

    2021

  3. [11]

    Class-balanced loss based on effective number of samples

    Cui, Y., Jia, M., Lin, T.-Y., Song, Y., and Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9268--9277, 2019

    2019

  4. [12]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 248--255, 2009. doi:10.1109/CVPR.2009.5206848

    2009

  5. [13]

    Robust learning with progressive data expansion against spurious correlation

    Deng, Y., Yang, Y., Mirzasoleiman, B., and Gu, Q. Robust learning with progressive data expansion against spurious correlation. In Advances in Neural Information Processing Systems, volume 36, 2024

    2024

  6. [14]

    A., and Brendel, W

    Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., and Brendel, W. Imagenet-trained CNN s are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019

    2019

  7. [15]

    Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., and Wichmann, F. A. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2 0 (11): 0 665--673, 2020

    2020

  8. [16]

    Umix: Improving importance weighting for subpopulation shift via uncertainty-aware mixup

    Han, Z., Liang, Z., Yang, F., Liu, L., Li, L., Bian, Y., Zhao, P., Wu, B., Zhang, C., and Yao, J. Umix: Improving importance weighting for subpopulation shift via uncertainty-aware mixup. In Advances in Neural Information Processing Systems, volume 35, pp.\ 37704--37718, 2022

    2022

  9. [17]

    and Garcia, E

    He, H. and Garcia, E. A. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21 0 (9): 0 1263--1284, 2009

    2009

  10. [18]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 770--778, 2016

    2016

  11. [19]

    Eva: Erasing spurious correlations with activations

    He, Q., Xu, K., and Yao, A. Eva: Erasing spurious correlations with activations. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  12. [20]

    Variable selection--a review and recommendations for the practicing statistician

    Heinze, G., Wallisch, C., and Dunkler, D. Variable selection--a review and recommendations for the practicing statistician. Biometrical journal, 60 0 (3): 0 431--449, 2018

    2018

  13. [21]

    Natural adversarial examples

    Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., and Song, D. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15262--15271, 2021

    2021

  14. [22]

    Izmailov, P., Kirichenko, P., Gruver, N., and Wilson, A. G. On feature learning in the presence of spurious correlations. In Advances in Neural Information Processing Systems, volume 35, pp.\ 38516--38532, 2022

    2022

  15. [23]

    Kenton, J. D. M.-W. C. and Toutanova, L. K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pp.\ 4171--4186, 2019

    2019

  16. [24]

    Learning debiased classifier with biased committee

    Kim, N., Hwang, S., Ahn, S., Park, J., and Kwak, S. Learning debiased classifier with biased committee. In Advances in Neural Information Processing Systems, volume 35, pp.\ 18403--18415, 2022

    2022

  17. [25]

    Kirichenko, P., Izmailov, P., and Wilson, A. G. Last layer re-training is sufficient for robustness to spurious correlations. In International Conference on Learning Representations, 2023

    2023

  18. [26]

    W., Sagawa, S., Marklund, H., Xie, S

    Koh, P. W., Sagawa, S., Marklund, H., Xie, S. M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R. L., Gao, I., et al. Wilds: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning, pp.\ 5637--5664. PMLR, 2021

    2021

  19. [27]

    Towards last-layer retraining for group robustness with fewer annotations

    LaBonte, T., Muthukumar, V., and Kumar, A. Towards last-layer retraining for group robustness with fewer annotations. In Advances in Neural Information Processing Systems, volume 36, 2024

    2024

  20. [28]

    Bias amplification enhances minority group performance

    Li, G., Liu, J., and Hu, W. Bias amplification enhances minority group performance. Transactions on Machine Learning Research, 2024

    2024

  21. [29]

    Z., Haghgoo, B., Chen, A

    Liu, E. Z., Haghgoo, B., Chen, A. S., Raghunathan, A., Koh, P. W., Sagawa, S., Liang, P., and Finn, C. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pp.\ 6781--6792. PMLR, 2021

    2021

  22. [30]

    Deep learning face attributes in the wild

    Liu, Z., Luo, P., Wang, X., and Tang, X. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 3730--3738, 2015

    2015

  23. [31]

    Learning from failure: De-biasing classifier from biased classifier

    Nam, J., Cha, H., Ahn, S., Lee, J., and Shin, J. Learning from failure: De-biasing classifier from biased classifier. In Advances in Neural Information Processing Systems, volume 33, pp.\ 20673--20684, 2020

    2020

  24. [32]

    Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation

    Nam, J., Kim, J., Lee, J., and Shin, J. Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation. In International Conference on Learning Representations, 2022

    2022

  25. [33]

    Qiu, S., Potapczynski, A., Izmailov, P., and Wilson, A. G. Simple and fast group robustness by automatic feature reweighting. In International Conference on Machine Learning, pp.\ 28448--28467. PMLR, 2023

    2023

  26. [34]

    W., Hashimoto, T

    Sagawa, S., Koh, P. W., Hashimoto, T. B., and Liang, P. Distributionally robust neural networks. In International Conference on Learning Representations, 2019

    2019

  27. [35]

    W., and Liang, P

    Sagawa, S., Raghunathan, A., Koh, P. W., and Liang, P. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pp.\ 8346--8356. PMLR, 2020

    2020

  28. [36]

    Unsupervised learning of debiased representations with pseudo-attributes

    Seo, S., Lee, J.-Y., and Han, B. Unsupervised learning of debiased representations with pseudo-attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16742--16751, 2022

    2022

  29. [37]

    and Feizi, S

    Singla, S. and Feizi, S. Salient imagenet: How to discover spurious features in deep learning? In International Conference on Learning Representations, 2021

    2021

  30. [38]

    No subclass left behind: Fine-grained robustness in coarse-grained classification problems

    Sohoni, N., Dunnmon, J., Angus, G., Gu, A., and R \'e , C. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. In Advances in Neural Information Processing Systems, volume 33, pp.\ 19339--19352, 2020

    2020

  31. [39]

    A., and Grangetto, M

    Tartaglione, E., Barbano, C. A., and Grangetto, M. End: Entangling and disentangling deep representations for bias correction. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pp.\ 13508--13517, 2021

    2021

  32. [40]

    and Shenoy, P

    Tiwari, R. and Shenoy, P. Overcoming simplicity bias in deep networks using a feature sieve. In International Conference on Machine Learning, pp.\ 34330--34343. PMLR, 2023

    2023

  33. [41]

    Vapnik, V. N. An overview of statistical learning theory. IEEE Transactions on Neural Networks, 10 0 (5): 0 988--999, 1999

    1999

  34. [42]

    Causal attention for unbiased visual recognition

    Wang, T., Zhou, C., Sun, Q., and Zhang, H. Causal attention for unbiased visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 3091--3100, 2021

    2021

  35. [43]

    Navigate beyond shortcuts: Debiased learning through the lens of neural collapse

    Wang, Y., Sun, J., Wang, C., Zhang, M., and Yang, M. Navigate beyond shortcuts: Debiased learning through the lens of neural collapse. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12322--12331, 2024

    2024

  36. [44]

    Caltech-UCSD Birds 200

    Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., and Perona, P. Caltech-UCSD Birds 200 . Technical Report CNS-TR-2010-001, California Institute of Technology, 2010

    2010

  37. [45]

    A broad-coverage challenge corpus for sentence understanding through inference

    Williams, A., Nangia, N., and Bowman, S. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1...

    2018

  38. [46]

    Discover and cure: Concept-aware mitigation of spurious correlation

    Wu, S., Yuksekgonul, M., Zhang, L., and Zou, J. Discover and cure: Concept-aware mitigation of spurious correlation. In International Conference on Machine Learning, pp.\ 37765--37786. PMLR, 2023

    2023

  39. [47]

    Y., Engstrom, L., Ilyas, A., and Madry, A

    Xiao, K. Y., Engstrom, L., Ilyas, A., and Madry, A. Noise or signal: The role of image backgrounds in object recognition. In International Conference on Learning Representations, 2021

    2021

  40. [48]

    Change is hard: a closer look at subpopulation shift

    Yang, Y., Zhang, H., Katabi, D., and Ghassemi, M. Change is hard: a closer look at subpopulation shift. In International Conference on Machine Learning, pp.\ 39584--39622, 2023

    2023

  41. [49]

    Improving out-of-distribution robustness via selective augmentation

    Yao, H., Wang, Y., Li, S., Zhang, L., Liang, W., Zou, J., and Finn, C. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning, pp.\ 25407--25437. PMLR, 2022

    2022

  42. [50]

    Freeze then train: Towards provable representation learning under spurious correlations and feature noise

    Ye, H., Zou, J., and Zhang, L. Freeze then train: Towards provable representation learning under spurious correlations and feature noise. In International Conference on Artificial Intelligence and Statistics, pp.\ 8968--8990. PMLR, 2023

    2023

  43. [51]

    Spurious correlations in machine learning: A survey

    Ye, W., Zheng, G., Cao, X., Ma, Y., and Zhang, A. Spurious correlations in machine learning: A survey. ICML Workshop on Data-Centric Machine Learning Research, 2024

    2024

  44. [52]

    N., and Lopez-Paz, D

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

    2018

  45. [53]

    S., Zhang, H

    Zhang, M., Sohoni, N. S., Zhang, H. R., Finn, C., and Re, C. Correct- N - C ontrast: A contrastive approach for improving robustness to spurious correlations. In Proceedings of the 39th International Conference on Machine Learning, volume 162, pp.\ 26484--26516. PMLR, 2022

    2022

  46. [54]

    Poisoning for debiasing: Fair recognition via eliminating bias uncovered in data poisoning

    Zhang, Y., Wang, Z., Hu, R., Duan, X., ZHENG, Y., Huai, B., Han, J., and Sang, J. Poisoning for debiasing: Fair recognition via eliminating bias uncovered in data poisoning. In ACM Multimedia, 2024

    2024

  47. [55]

    Learning debiased representations via conditional attribute interpolation

    Zhang, Y.-K., Wang, Q.-W., Zhan, D.-C., and Ye, H.-J. Learning debiased representations via conditional attribute interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7599--7608, 2023

    2023

  48. [56]

    Benchmarking spurious bias in few-shot image classifiers

    Zheng, G., Ye, W., and Zhang, A. Benchmarking spurious bias in few-shot image classifiers. In European Conference on Computer Vision, pp.\ 346--364. Springer, 2024 a

    2024

  49. [57]

    Learning robust classifiers with self-guided spurious correlation mitigation

    Zheng, G., Ye, W., and Zhang, A. Learning robust classifiers with self-guided spurious correlation mitigation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pp.\ 5599--5607, 2024 b

    2024

  50. [58]

    Spuriousness-aware meta-learning for learning robust classifiers

    Zheng, G., Ye, W., and Zhang, A. Spuriousness-aware meta-learning for learning robust classifiers. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 4524--4535, 2024 c

    2024

  51. [59]

    Shortcutprobe: Probing prediction shortcuts for learning robust models

    Zheng, G., Ye, W., and Zhang, A. Shortcutprobe: Probing prediction shortcuts for learning robust models. In Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, 2025

    2025

  52. [60]

    Places: A 10 million image database for scene recognition

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., and Torralba, A. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40 0 (6): 0 1452--1464, 2017

    2017

Pith tools

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