Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Elastic Representation: Mitigating Spurious Correlations for Group Robustness

T0 review · 3 major / 6 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 A representation regularizer that pairs nuclear and Frobenius norms lifts worst-group accuracy across three benchmarks without needing group labels.

desk verdict Empirically useful regularizer, but the theory section analyzes a different objective and the 'minimum sacrifice' claim is contradicted by the CivilComments numbers. read the letter →

arxiv 2502.09850 v1 pith:BLSGIBUJ submitted 2025-02-14 cs.LG

Tao Wen , Zihan Wang , Quan Zhang , Qi Lei This is my paper · ORCID
classification cs.LG
keywords spuriouscorrelationsgrouprobustnessworst-groupaccuracynuclearnormregularizationFrobeniusrepresentationlearningelasticnetanalogyout-of-distributiongeneralization
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 proposes a regularizer, Elastic Representation (ElRep), that adds a nuclear-norm penalty and a Frobenius-norm penalty to the last-layer feature representation of a neural network. The claim is that this two-part penalty weakens the model's reliance on spurious features, which correlate with the label only in training data, while keeping enough correlated invariant features to generalize to minority groups. ElRep needs no group annotations, so it can be dropped into standard training or combined with methods that already use group information. On CelebA, Waterbirds, and CivilComments, adding ElRep improves worst-group accuracy across ERM, upweighting, subsampling, GroupDRO, and PDE, and the paper argues the average in-distribution accuracy is hardly harmed. A reader should care because group robustness usually costs either extra annotations or a noticeable drop in overall performance; ElRep claims to avoid both.

What carries the argument

The engine is the objective $\min_{W,\Phi} L(W,\Phi) + \lambda_1 \|\Phi(x)\|_* + \lambda_2 \|\Phi(x)\|_F^2$, where the nuclear norm $\|A\|_* = \mathrm{Tr}((A^\top A)^{1/2})$ sums the singular values and the Frobenius norm $\|A\|_F = (\mathrm{Tr}(A^\top A))^{1/2}$ is the root-sum-of-squared singular values. The nuclear term selects a sparse set of feature directions, and the Frobenius term keeps the retained directions diverse, analogously to the elastic net's $\ell_1+\ell_2$ combination. The proofs work in a linear-regression surrogate where ElRep becomes $\lambda_1\|\theta\|_1 + \lambda_2\|\theta\|_2$ on the regression weights, and Proposition 5.2 gives explicit thresholds showing that the combined penalty keeps two positively correlated features in cases where $\ell_1$ alone would zero one of them.

What would settle it

Train a linear model on data with two strongly correlated features, one invariant and one spurious, under the actual representation-level objective of Equation (1), and measure which features survive; if the representation-level penalty drops the invariant feature in the same regime where the weight-level $\ell_1+\ell_2$ theory says it should be kept, the theory does not carry over to the proposed method.

Watch

Extended reading notes

Core claim

The central discovery is that penalizing the representation with the nuclear norm and the Frobenius norm simultaneously, rather than either alone, mitigates spurious correlations without group information. The nuclear norm (the sum of the singular values of the representation matrix) pushes the representation toward low rank, pruning spurious feature directions, while the Frobenius norm (the root-sum-of-squared singular values) counteracts over-sparsification, preserving correlated invariant features that nuclear norm alone would discard. The paper demonstrates empirically that ElRep raises worst-group accuracy on CelebA, Waterbirds, and CivilComments, and that combining it with existing group-robust methods improves their worst-group accuracy as well. The theoretical section argues that the regularization does not hurt in-distribution prediction and that the $\ell^1$-plus-$\ell^2$ style penalty keeps correlated features more often than $\ell^1$ alone.

Load-bearing premise

The theoretical results are proven for a simplified version in which the penalties are applied to the classifier's weights, not to the learned representation itself, and the paper asserts these are similar without showing the two formulations are equivalent.

Editorial extensions

If this is right

  • Adding ElRep to ERM is a group-agnostic route to better worst-group accuracy, with the average accuracy loss typically small and sometimes negative on image benchmarks.
  • Because ElRep does not require group labels, it can be added to any model where group annotations are impractical, including two-stage train-twice pipelines.
  • The nuclear-plus-Frobenius combination addresses a known failure mode of nuclear norm alone, where highly correlated invariant features are dropped and fail to reappear under test-time correlation shifts.
  • Combining ElRep with group-aware methods such as GroupDRO, PDE, subsampling, and upweighting gives further worst-group gains, with the largest gains appear on the simplest baselines such as ERM.
  • The in-distribution preservation bound implies that large-sample training with ElRep has prediction risk close to the ground-truth model, so practitioners do not have to choose between overall accuracy and minority-group robustness.

Reading between the lines

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

  • The paper's theory analyzes $\ell_1+\ell_2$ penalties on regression weights, not nuclear-plus-Frobenius penalties on the representation itself; a rigorous derivation of that equivalence is needed before those results can be treated as applying to the actual ElRep objective.
  • Because ElRep acts on the representation rather than the classifier, its benefit should grow when the feature set is large and redundant, making fine-grained classification with many correlated visual attributes a natural test bed.
  • ElRep's mechanism suggests a cheap robustness audit: run the same model with and without the two penalties and compare worst-group accuracy, since large gaps indicate spurious reliance.
  • It could be combined with automatic minority-group discovery methods, potentially replacing group annotations entirely while retaining the signal that group-aware methods use.
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 / 6 minor

Summary. The paper proposes Elastic Representation (ElRep), a regularizer that adds nuclear-norm and Frobenius-norm penalties to the last-layer representation of a neural network, with the aim of reducing reliance on spurious correlations without using group annotations. The method is evaluated on CelebA, Waterbirds, and CivilComments, where it is combined with ERM, upweighting, subsampling, GroupDRO, and PDE; the reported worst-group accuracies generally improve relative to the corresponding baselines. The paper also includes a synthetic experiment and a theoretical section claiming that ElRep has minimal negative impact on in-distribution performance and that the added Frobenius penalty helps retain correlated invariant features. The main technical weakness is that the theoretical analysis in Section 5 analyzes a different objective (l1 plus l2 penalization of the regression weight vector) rather than the nuclear-plus-Frobenius representation penalty in Eq. (1), and the claimed equivalence is not derived.

Significance. If the empirical findings are taken at face value, ElRep is a simple, group-annotation-free regularizer that can be plugged into several existing group-robustness methods and yields consistent worst-group accuracy improvements on standard benchmarks, which would be practically useful. The authors also make their code available and use established evaluation protocols. However, the theoretical claims in the abstract and Section 5 outrun the analysis: Theorem 5.1 and Proposition 5.2 concern an elastic net on weights, not the proposed representation regularizer, and the abstract's assertion of 'minimum negative impacts' is directly contradicted by the large average-accuracy drops on CivilComments in Table 3. The paper's contribution is therefore best assessed as an empirical method with a stated but unsupported theoretical justification.

major comments (3)
  1. [Section 5, Eq. (2)] The theoretical analysis does not analyze the proposed objective in Eq. (1). Section 5 replaces it with Eq. (2), which penalizes the regression weight theta with l1 and l2 norms, and states that this 'has a similar effect' to penalizing the representation. No derivation or formal equivalence is provided. In the linear model f(x) = theta^T x, the representation penalty in Eq. (1) would act on Phi(x), and because Phi and W interact in the risk L(W, Phi), the objective is not generally reducible to a penalty on theta alone. Furthermore, if Phi(x) is a vector, then the nuclear norm of Phi(x) equals its Frobenius norm, and the two penalties in Eq. (1) coincide, so the claimed elastic-net analogy collapses; if Phi(X) is a batch matrix, the theory must be restated in terms of the singular values of that matrix. Theorem 5.1 and Proposition 5.2 therefore do not establish the stated properties of ElRep as defined in Eq. (1).
  2. [Section 5.1, Theorem 5.1 and Abstract] The claim that ElRep has 'minimum negative impacts' on in-distribution predictions is not supported by Theorem 5.1, which only gives an upper bound on the excess risk for the elastic-net-on-weights estimator and makes no comparison with other regularizers or with the unregularized estimator. Moreover, the claim is empirically contradicted by Table 3: on CivilComments, adding ElRep to GroupDRO reduces average accuracy from 89.6 to 79.0, and adding it to PDE reduces average accuracy from 86.3 to 80.7. These are substantial degradations, not minimal impacts. The wording in the abstract and conclusion should be revised to match the actual evidence.
  3. [Section 4.1, Synthetic Data] The synthetic experiment states that 'the nuclear- and Frobenius-norms are reduced to l1- and l2-norms of a, respectively,' but this reduction is not derived. The representation Phi(x) is described as a (k+2) by d matrix with rows involving elementwise products a_i and z_i; for a general matrix of this form, the nuclear norm is not equal to the l1 norm of the coefficient vector a, nor is the Frobenius norm equal to the l2 norm of a, unless strong orthogonality and scaling conditions on the z_i are imposed. Without such a derivation, the synthetic results in Tables 1 and 2 do not demonstrate the behavior of the actual ElRep objective in Eq. (1).
minor comments (6)
  1. [Section 5.1, Theorem 5.1] The theorem statement says 'choose proper lambda1, lambda2, lambda3' but Eq. (2) contains only lambda1 and lambda2; lambda3 is never defined and appears to be a typographical leftover.
  2. [Appendix B, Proof of Lemma B.1] There is a typo in the definition of A: 'X_T^T Z_t' should read 'X_t^T Z_t', and 'with probability at list 1 - delta' should read 'at least.'
  3. [Section 3.3] The statement that regularizing the representation is 'a dual problem' to regularizing the weight W is imprecise; the notion of duality is not defined, and Figure 2 is not described in the text. This analogy should either be stated formally or presented as informal intuition.
  4. [Section 5, end of 5.1] The sentence 'The analysis of OOD performance is not included because more assumptions of the testing domain are needed' is important and should be reflected in the abstract and introduction, since the title and opening claims concern mitigating spurious correlations and improving group robustness.
  5. [Table 4] In the UW rows for CivilComments, a stray 'N/A' appears in the UW(FN) row, while the adjacent NN and Ours rows are empty; the formatting should be made consistent.
  6. [Introduction, Contribution 3] The claim that ElRep is 'without extra computational cost' is questionable: computing the nuclear norm of a batch of representations requires a singular value decomposition at each training step, which is an additional computational burden relative to plain ERM.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction found: the empirical claims rest on external benchmarks, and the main theory caveat (Section 5 analyzes an l1+l2 weight penalty rather than the representation penalty) is an unsupported analogy, not a fit-to-target loop.

full rationale

The empirical core of the paper is not circular. ElRep is added to ERM, UW, Subsample, GroupDRO, and PDE, and the resulting worst-group and average accuracies are compared against published, external benchmarks on CelebA, Waterbirds, and CivilComments; the two regularization strengths are selected by cross-validation, so no fitted parameter is later relabeled as a prediction. The only self-referential element is in the theoretical proof: Theorem 5.1 relies on 'Lemma C.10 in (Du et al., 2021)' (a paper co-authored by Qi Lei) to convert training excess risk into population excess risk, and the text also remarks that for nuclear-norm regularization 'the bound only differs in constant coefficients according to Du et al. (2021).' This citation supplies a general subgaussian comparison lemma, not the paper's own target conclusion, so it is not load-bearing in a circularity sense. A separate, non-circular correctness caveat is that Section 5.1 introduces Eq. (2), which penalizes the regression weight theta with lambda1||theta||_1 + lambda2||theta||_2, while the proposed ElRep objective in Eq. (1) penalizes the representation norms lambda1||Phi(x)||_* + lambda2||Phi(x)||_F^2; the text bridges these two only with the phrase 'similar effect.' Theorem 5.1 is therefore best read as a consistency bound for l1+l2-regularized linear regression rather than a direct theorem about the representation penalty, but this is an unsupported analogy, not an input-output equivalence. Overall, the central empirical claims have independent benchmark content, so the circularity score is low.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The paper introduces no new physical entity. The main unstated burden is the equivalence between representation norms and weight norms, which all of the theoretical claims in Section 5 rely on. The two regularization hyperparameters λ1 and λ2 are fitted to validation data, and the synthetic experiment adds several hand-chosen coefficients. The remaining axioms are standard distributional assumptions or accepted background results.

free parameters (3)
  • nuclear norm coefficient λ1 = selected by cross-validation in range 1e-4 to 1e-3; exact values not reported
    Controls the nuclear-norm penalty in Equation (1); chosen separately for each dataset and baseline using validation accuracy.
  • Frobenius norm coefficient λ2 = selected by cross-validation in range 1e-5 to 1e-4; exact values not reported
    Controls the Frobenius-norm penalty in Equation (1); chosen together with λ1 by cross-validation.
  • synthetic data-generation coefficients and dimensions = k=3, d=100, n=120, ρ=0.5
    Chosen by hand for the synthetic experiment in Section 4.1; these settings do not affect the real-data claims.
assumptions (7)
  • domain assumption Assumption 5.1: there exists a positive semidefinite matrix Σ such that Σ_t ⪯ Σ for every training domain t.
    Used in Lemma B.1 and Theorem 5.1 to control the cross term and population excess risk; a standard bounded-covariance condition.
  • domain assumption Assumption 5.2: the whitened feature vector x̄ is ρ²-subgaussian for every training domain.
    Used in the proof of Theorem 5.1 through Du et al. Lemma C.10 to convert the empirical training bound into a population excess risk bound.
  • ad hoc to paper Penalizing the representation Φ(x) with nuclear and Frobenius norms is equivalent in effect to penalizing the weight vector θ with l1 and l2 norms.
    Section 5 introduces Equation (2) as the objective for the same problem and states that l1/l2 on θ 'has a similar effect' to penalizing the representation, but no derivation is provided. The theoretical results are proved for Equation (2), not for Equation (1).
  • domain assumption The data-generation process decomposes features into invariant, nuanced, and spurious components.
    This decomposition underlies the synthetic experiment and the conceptual motivation in Sections 1 and 4.1. It is a common assumption in the spurious-correlation literature rather than a result proved here.
  • domain assumption Standard ERM models learn both spurious and invariant features, so the linear classifier is responsible for the minority-group failure.
    Borrowed from Kirichenko et al. (2023) and Izmailov et al. (2022); this motivates regularizing the representation instead of retraining the classifier.
  • standard math Lemma C.10 of Du et al. (2021) relating empirical and population risk under subgaussian features.
    Used in the proof of Theorem 5.1 without restating the lemma; accepted as a prior result.
  • standard math Matrix concentration inequalities from Tropp et al. (2015), including Theorem 7.3.1.
    Used in Lemma B.1 to bound the random cross term; standard background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Elastic Representation: Mitigating Spurious Correlations for Group Robustness." pith.science (2026). https://pith.science/paper/BLSGIBUJ

@misc{pith2026250209850,
  author       = {Pith},
  title        = {Pith review of: Elastic Representation: Mitigating Spurious Correlations for Group Robustness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLSGIBUJ}},
  note         = {Machine review of arXiv:2502.09850}
}
read the original abstract

Deep learning models can suffer from severe performance degradation when relying on spurious correlations between input features and labels, making the models perform well on training data but have poor prediction accuracy for minority groups. This problem arises especially when training data are limited or imbalanced. While most prior work focuses on learning invariant features (with consistent correlations to y), it overlooks the potential harm of spurious correlations between features. We hereby propose Elastic Representation (ElRep) to learn features by imposing Nuclear- and Frobenius-norm penalties on the representation from the last layer of a neural network. Similar to the elastic net, ElRep enjoys the benefits of learning important features without losing feature diversity. The proposed method is simple yet effective. It can be integrated into many deep learning approaches to mitigate spurious correlations and improve group robustness. Moreover, we theoretically show that ElRep has minimum negative impacts on in-distribution predictions. This is a remarkable advantage over approaches that prioritize minority groups at the cost of overall performance.

Figures

Figures reproduced from arXiv: 2502.09850 by the authors.

Figure 1
Figure 1. A long-tailed Jaeger, a waterbird on a land background, from the waterbirds dataset (Sagawa et al., 2019). The heat maps depict the pixel contri￾butions to bird type prediction using Grad-CAM (Sel￾varaju et al., 2019). From left to right are the origi￾nal image, ERM, ERM with nuclear norm, and ERM with nuclear and Frobenius norms, respectively. ERM learns features including background areas. ERM with nuclear norm fo… view at source ↗
Figure 2
Figure 2. Connections between ElRep and elastic net. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. This is attributed to the theoretical under [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Left: The difference in the worst-group accu [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 4
Figure 4. Figure 4: Accuracy per group and average accuracy against the log of λ1 (left) and λ2 (right). As their value increases, the accuracy of the two minority groups will gradually increase and eventually surpass the average accuracy. The trend is reversed for the two majority groups…
Figure 5
Figure 5. Figure 5: The two majority groups downsampled to about 1%. Reversed trends are observed. useful in cases where we only care about small group accuracy since we can set arbitrarily large values for λ1 and λ2, as long as the regularization term does not overwhelm the classificatio…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Bridging Distribution Shift and AI Safety: Conceptual and Methodological Synergies

    cs.LG 2025-05 conditional novelty 5.0 of 10

    The paper proposes a one-to-one mapping between six causes of distribution shift and several AI safety issues, arguing for mutual method transfer through aligned definitions.

Reference graph

Works this paper leans on

45 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Chen, Y., Huang, W., Zhou, K., Bian, Y., Han, B., and Cheng, J. (2023). Understanding and improving feature learning for out-of-distribution generalization. Advances in Neural Information Processing Systems , 36

    work page 2023

  2. [2]

    Deng, Y., Yang, Y., Mirzasoleiman, B., and Gu, Q. (2023). Robust learning with progressive data expansion against spurious correlation. Advances in neural information processing systems , 36

    work page 2023

  3. [3]

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2019). Bert: Pre-training of deep bidirectional transformers for language understanding

    2019

  4. [4]

    S., Hu, W., Kakade, S

    Du, S. S., Hu, W., Kakade, S. M., Lee, J. D., and Lei, Q. (2021). Few-shot learning via learning the representation, provably. In 9th International Conference on Learning Representations, ICLR 2021

    work page 2021

  5. [5]

    Du, Y., Yan, J., Chen, Y., Liu, J., Zhao, S., She, Q., Wu, H., Wang, H., and Qin, B. (2023). Less learn shortcut: Analyzing and mitigating learning of spurious feature-label correlation. IJCAI

    work page 2023

  6. [6]

    u gelgen, J., and Sch \

    Eastwood, C., Singh, S., Nicolicioiu, A. L., Vlastelica Pogan c i \'c , M., von K \"u gelgen, J., and Sch \"o lkopf, B. (2023). Spuriosity didn’t kill the classifier: Using invariant predictions to harness spurious features. Advances in Neural Information Processing Systems , 36

    work page 2023

  7. [7]

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

    2020

  8. [8]

    A., and Brendel, W

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

    work page 2019

Show all 45 references
  1. [9]

    Goel, K., Gu, A., Li, Y., and R \'e , C. (2020). Model patching: Closing the subgroup performance gap with data augmentation. arXiv preprint arXiv:2008.06775

    2020 arXiv

  2. [10]

    Haghtalab, N., Jordan, M., and Zhao, E. (2022). On-demand sampling: Learning optimally from multiple distributions. Advances in Neural Information Processing Systems , 35:406--419

    2022

  3. [11]

    He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep residual learning for image recognition

    2015

  4. [12]

    S., Langlotz, C

    Huang, S.-C., Chaudhari, A. S., Langlotz, C. P., Shah, N., Yeung, S., and Lungren, M. P. (2022). Developing medical imaging ai for emerging infectious diseases. nature communications , 13(1):7060

    2022

  5. [13]

    Y., Arjovsky, M., Pezeshki, M., and Lopez-Paz, D

    Idrissi, B. Y., Arjovsky, M., Pezeshki, M., and Lopez-Paz, D. (2022). Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning , pages 336--351. PMLR

    2022

  6. [14]

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

    2022

  7. [15]

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

    2023

  8. [16]

    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., Lee, T., David, E., Stavness, I., Guo, W., Earnshaw, B. A., Haque, I. S., Beery, S., Leskovec, J., Kundaje, A., Pierson, E., Levine, S., Finn, C., and...

    2021

  9. [17]

    LaBonte, T., Muthukumar, V., and Kumar, A. (2023). Towards last-layer retraining for group robustness with fewer annotations. Advances in Neural Information Processing Systems , 36

    2023

  10. [18]

    C., and Sidford, A

    Levy, D., Carmon, Y., Duchi, J. C., and Sidford, A. (2020). Large-scale methods for distributionally robust optimization. Advances in Neural Information Processing Systems , 33:8847--8860

    2020

  11. [19]

    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. (2021). Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning , pages 6781--6792. PMLR

    2021

  12. [20]

    Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learning face attributes in the wild

    2015

  13. [21]

    M., and Sch \"o lkopf, B

    Lu, C., Wu, Y., Hern \'a ndez-Lobato, J. M., and Sch \"o lkopf, B. (2021). Nonlinear invariant risk minimization: A causal approach. arXiv preprint arXiv:2102.12353

    2021 arXiv

  14. [22]

    Moayeri, M., Wang, W., Singla, S., and Feizi, S. (2023). Spuriosity rankings: sorting data to measure and mitigate biases. Advances in Neural Information Processing Systems , 36:41572--41600

    2023

  15. [23]

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

    2020

  16. [24]

    G., and Lei, Q

    Phan, H., Wilson, A. G., and Lei, Q. (2024). Controllable prompt tuning for balancing group distributional robustness

    2024

  17. [25]

    Piratla, V., Netrapalli, P., and Sarawagi, S. (2022). Focus on the common good: Group distributional robustness follows

    2022

  18. [26]

    Rosenfeld, E., Ravikumar, P., and Risteski, A. (2022). Domain-adjusted regression or: Erm may already learn features sufficient for out-of-distribution generalization. arXiv preprint arXiv:2202.06856

    2022 arXiv

  19. [27]

    W., Hashimoto, T

    Sagawa, S., Koh, P. W., Hashimoto, T. B., and Liang, P. (2019). Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. International Conference on Learning Representations

    2019

  20. [28]

    W., and Liang, P

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

    2020

  21. [29]

    R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D

    Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D. (2019). Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision , 128(2):336–359

    2019

  22. [30]

    Shah, H., Tamuly, K., Raghunathan, A., Jain, P., and Netrapalli, P. (2020). The pitfalls of simplicity bias in neural networks. Advances in Neural Information Processing Systems , 33:9573--9585

    2020

  23. [31]

    Shi, Z., Ming, Y., Fan, Y., Sala, F., and Liang, Y. (2024). Domain generalization via nuclear norm regularization. In Conference on Parsimony and Learning , pages 179--201. PMLR

    2024

  24. [32]

    Sun, X., Wu, B., Zheng, X., Liu, C., Chen, W., Qin, T., and Liu, T.-Y. (2021). Recovering latent causal factor for generalization to distributional shifts. Advances in Neural Information Processing Systems , 34:16846--16859

    2021

  25. [33]

    Tropp, J. A. et al. (2015). An introduction to matrix concentration inequalities. Foundations and Trends in Machine Learning , 8(1-2):1--230

    2015

  26. [34]

    Veitch, V., D'Amour, A., Yadlowsky, S., and Eisenstein, J. (2021). Counterfactual invariance to spurious correlations in text classification. Advances in neural information processing systems , 34:16196--16208

    2021

  27. [35]

    Wah, C., Branson, S., Welinder, P., Perona, P., and Belongie, S. (2011). The Caltech-UCSD Birds-200-2011 Dataset

    2011

  28. [36]

    Wu, S., Yuksekgonul, M., Zhang, L., and Zou, J. (2023). Discover and cure: concept-aware mitigation of spurious correlation. In Proceedings of the 40th International Conference on Machine Learning , ICML'23. JMLR.org

    2023

  29. [37]

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

    2021

  30. [38]

    K., and Mirzasoleiman, B

    Yang, Y., Gan, E., Dziugaite, G. K., and Mirzasoleiman, B. (2024). Identifying spurious biases early in training through the lens of simplicity bias. In International Conference on Artificial Intelligence and Statistics , pages 2953--2961. PMLR

    2024

  31. [39]

    Ye, W., Zheng, G., Cao, X., Ma, Y., Hu, X., and Zhang, A. (2024). Spurious correlations in machine learning: A survey. arXiv preprint arXiv:2402.12715

    2024

  32. [40]

    Yenamandra, S., Ramesh, P., Prabhu, V., and Hoffman, J. (2023). Facts: First amplify correlations and then slice to discover bias. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4794--4804

    2023

  33. [41]

    Zhang, J., Menon, A., Veit, A., Bhojanapalli, S., Kumar, S., and Sra, S. (2021). Coping with label shift via distributionally robust optimisation. International Conference on Learning Representations

    2021

  34. [42]

    S., Zhang, H

    Zhang, M., Sohoni, N. S., Zhang, H. R., Finn, C., and R \'e , C. (2022). Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. In International Conference on Machine Learning . PMLR

    2022

  35. [43]

    S., Pan, X., and Lei, Q

    Zhong, Z. S., Pan, X., and Lei, Q. (2024). Bridging domains with approximately shared features. arXiv preprint arXiv:2403.06424

    2024 arXiv

  36. [44]

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., and Torralba, A. (2018). Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence , 40(6):1452--1464

    2018

  37. [45]

    and Hastie, T

    Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B: Statistical Methodology , 67(2):301--320

    2005

Pith tools

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