Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Bias as a Virtue: Rethinking Generalization under Distribution Shifts

T0 review · 4 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 Under a mean-shift model, the paper proves that in-distribution and out-of-distribution errors are negatively correlated when candidate-model bias spread is small relative to the shift, and gives a training-sequence selection method that…

desk verdict Neat theoretical observation buried under overclaiming and weak empirical support; the ADB framework is a real idea but this paper doesn't substantiate it. read the letter →

arxiv 2506.00407 v1 pith:CRAAJNME submitted 2025-05-31 cs.LG cs.AIstat.ML

Ruixuan Chen , Wentao Li , Jiahui Xiao , Yuchen Li , Yimin Tang , Xiaonan Wang This is my paper
classification cs.LGcs.AIstat.ML
keywords out-of-distributiongeneralizationmodelselectiondistributionshiftbias-variancetradeoffoptimaltransporttrainingsequencediversitycross-validationhalf-normal
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 argues that conventional cross-validation can select the wrong model under large distribution shifts. It proves that when candidate models' bias values $b$ follow a half-normal distribution with scale $k$, the in-distribution error $T(b)=b^2$ and the out-of-distribution error $U(b)=(b-\Delta)^2$ have negative Pearson correlation exactly when $k<\sqrt{2/\pi}\,\Delta$. In that regime, a model with higher in-distribution error is the better OOD choice. The paper's Adaptive Distribution Bridge (ADB) groups training sequences by their debiased Sinkhorn distance from the global training distribution and selects the highest-ID-error model from its Medium and High deviation groups, reporting up to 26.8% lower OOD MAE than 10-fold cross-validation.

What carries the argument

The load-bearing object is the correlation identity between squared error components under a half-normal bias population. The paper derives $\operatorname{Cov}(T,U)=2k^3(k-\alpha\Delta)$ from the half-normal moment generating function, so the sign of the ID-OOD correlation is governed by $k-\alpha\Delta$. ADB operationalizes $k$ by classifying random training permutations into Low, Medium, and High deviation groups using debiased Sinkhorn optimal-transport distances from the global training distribution, then training models on the more diverse sequences.

What would settle it

Train many models under ADB's deviation groups on a dataset with a label-space Wasserstein shift above 0.6, record each model's ID and OOD errors, and fit the empirical distribution of the bias parameter $b$. If the sample fourth moment departs from $3k^4$, if the fitted $k$ exceeds $\sqrt{2/\pi}\Delta$, or if the empirical Pearson correlation between ID and OOD error is not negative in that regime, the paper's central claim is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that, under distribution shift, higher in-distribution bias can be a virtue rather than a defect: if the spread of bias across candidate models is small relative to the shift, the errors measured in-distribution and out-of-distribution are negatively correlated, so minimizing ID error selects worse OOD models. The paper states the result as a theorem in a mean-shift model: with $T(b)=b^2$, $U(b)=(b-\Delta)^2$, and $b$ half-normal with scale $k$, $\rho_{T,U}<0$ iff $k<\alpha\Delta$ with $\alpha=\sqrt{2/\pi}$. It then presents ADB as the practical method that induces bias variation by ranking training sequences according to their distributional deviation, and reports that choosing the highest-ID-error model from Medium/High deviation sequences beats cross-validation.

Load-bearing premise

The whole selection rule depends on two unmeasured premises: that candidate bias values follow a half-normal distribution with a single scale $k$, and that ADB's deviation groups actually correspond to that $k$.

Editorial extensions

If this is right

  • Under significant distribution shift, standard cross-validation can actively choose the worst OOD model, because it targets the positive-correlation regime.
  • Any model-selection procedure that only minimizes ID error should be considered unsafe when the shift is large relative to bias diversity.
  • ADB's deviation grouping provides a target-distribution-free way to enter the negative-correlation regime, needing only the training data and an awareness that a shift exists.
  • The reported percentile ranks (above 74.4% and up to 91.2%) imply that ADB's selected strategies dominate most random training-sequence strategies, not just the CV baseline.
  • Because Cumulative deviation evaluation outperforms Batchwise on both datasets, practices that account for full training history should be preferred when compute allows.

Reading between the lines

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

  • A testable extension would estimate $k$ and $\Delta$ from an ID validation set plus one held-out shifted set; if the ratio $k/(\alpha\Delta)<1$ is confirmed before deployment, the higher-ID-error selection rule becomes a predictive tool rather than a post-hoc explanation.
  • The theory is stated for squared errors, while the headline empirical gains are in MAE; a direct check is needed to see whether the negative-correlation condition transfers to absolute-error losses.
  • The same half-normal framework suggests that the optimum is not maximal diversity but a calibrated one: the strongest negative correlation occurs at the smallest feasible $k$, so adding too much training-sequence diversity should eventually reverse the benefit.
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 / 4 minor

Summary. The paper argues that under distribution shift, models with higher in-distribution (ID) error can generalize better out-of-distribution (OOD). It presents a toy model in which ID error is T(b)=b^2, OOD error is U(b)=(b-Δ)^2, and b follows a half-normal distribution with scale k; Proposition 1 (Appendix A) shows the Pearson correlation between T and U is negative iff k<αΔ, α=√(2/π). The paper then introduces the Adaptive Distribution Bridge (ADB) framework, which classifies training permutations into Low/Medium/High deviation groups using debiased Sinkhorn distances, and reports that selecting the highest-ID-error model from the Medium/High groups yields OOD error reductions of up to 26.8% relative to cross-validation (Table 1).

Significance. If the empirical claims were valid, the paper would make a provocative contribution by questioning standard cross-validation practices under distribution shift. The theoretical derivation in Appendix A is transparent and correct as a calculus exercise, and the complexity analysis in Section 5.4 is a useful practical contribution. However, the empirical core is not currently sound: the reported CV baselines are inconsistent, the group-level results in Appendix B contradict the paper's own selection preference, and the proposed selection rule does not follow from the population-level negative-correlation result. Because the central empirical claims are load-bearing and unsupported, the manuscript does not currently meet the bar for publication.

major comments (4)
  1. [§5.2, Table 1] The cross-validation baseline for the same dataset is inconsistent across the two ADB variants. For T1S1 MAE, the Batchwise row reports CV=0.478±0.018 while the Cumulative row reports CV=0.514±0.020; for MSD MAE, Batchwise reports CV=0.868±0.005 while Cumulative reports CV=0.874±0.001. Since both rows compare against the same ``traditional cross-validation,'' the baseline should be numerically identical (or the difference must be explicitly explained as arising from separate experimental runs with different seeds). As stated, the reported improvement percentages (I) in Table 1 are not mutually consistent, and the paired t-tests in Table 7 cannot be interpreted without resolving this discrepancy.
  2. [§5.3 and Appendix B.3, Tables 5 and 6] The paper's claim that ADB ``favors Medium and High deviation groups'' (§5.3, and stated again in §5.2) is contradicted by the authors' own detailed results. In Table 6 (MSD, Cumulative), the Low deviation group has the best OOD performance (MAE 0.848, PR 95.0), while Medium (MAE 0.859, PR 71.3) and High (MAE 0.863, PR 62.5) are worse. In Table 5 (T1S1, Cumulative), the High group is substantially worse (MAE 0.436, PR 57.3) than both Low (MAE 0.403) and Medium (MAE 0.397). The empirical support for preferring Medium/High groups is therefore not present in the reported data.
  3. [§5.2 and §5.3] The selection rule--``sample 10 models proportionally from the Medium and High groups, select the one with the highest ID error, and report its OOD performance''--does not follow from a negative population correlation. Correlation is a statement about the joint distribution of T and U across models, but selecting the maximum-ID-error model is an extreme-order-statistic operation. In the paper's own quadratic model, OOD error U(b)=(b-Δ)^2 is minimized at b=Δ and increases for b>Δ, so a model with maximal ID error can lie beyond the OOD optimum and have worse OOD performance. The paper provides no theoretical or empirical justification that the selected extreme model is on the decreasing branch of the relationship. Moreover, because the authors pick the model with highest ID error after seeing the OOD results, the reported gains may reflect post-hoc selection rather than a predictive rule; a comparison against random selection from the same deviation groups, or an out-of-sample procedure, is needed.
  4. [§4.1 and §6] The bridge between the ADB deviation groups and the theoretical parameter k is never established. Section 4.1 asserts that ADB ``indirectly influences k,'' but the paper neither measures k nor estimates Δ, nor checks whether the model-bias distribution across ADB groups is approximately half-normal with any scale. The manuscript's own Limitations paragraph (Section 6) defers ``the quantitative relationship between induced distributional deviation levels and the resulting bias-diversity tradeoff'' to future work. Without this bridge, Proposition 1 remains a self-contained toy-model result rather than an explanation of the empirical findings.
minor comments (4)
  1. [§1 and §5.1] The abstract and Section 5.1 say ``three datasets,'' but only T1S1 and MSD are presented; the third dataset is never identified.
  2. [§5.1 and §5.2] Section 5.1 states that 30 models are trained per deviation group and each uses 50 permutation sequences per epoch, while Section 5.2 states that 10 models are sampled proportionally from Medium and High groups for Table 1. The relationship between these two numbers should be clarified.
  3. [§5.3] The text refers to ``Figure 1'' for the ID-OOD relationship, but no figure is included in the manuscript; either the figure should be added or the reference removed.
  4. [Appendix B.4, Table 7] The description of the paired t-test is unclear: ``10-fold cross-validation with fixed random seeds, analyzing OOD errors across all folds'' does not specify which models are paired, and the baseline inconsistency in Table 1 calls into question the validity of the paired comparison.

Circularity Check

1 steps flagged · score 6.0 of 10

The empirical selection rule is fitted to OOD test data in Figure 1 and then reported as a prediction on the same test set; the theoretical result does not supply the rule.

  1. fitted input called prediction [Section 5.3 ('Distribution-Error Relationship Analysis'), text after Figure 1; cf. Section 5.2, last paragraph before Table 1]
    "Based on the negative ID-OOD correlation patterns observed in Figure 1, we focus on Medium and High deviation groups for model selection. By selecting the highest ID error model from these groups, ADB leverages this correlation pattern to improve OOD performance. ... For Table 1, we sample 10 models proportionally from the Medium and High groups, select the one with the highest ID error, and report its OOD performance."

    The selection rule 'choose the highest-ID-error model from Medium/High groups' is adopted after inspecting Figure 1, which plots out-of-distribution MAE against in-distribution MAE on the OOD test set. The same OOD test set is then used to report the improvement in Table 1 and to claim a 'robust negative correlation' as empirical validation. Thus the reported OOD benefit is not an independent prediction; the rule was fitted to the test-set relationship and then re-reported as evidence for that relationship. The theory in Proposition 1 does not entail the rule, since a negative population correlation does not imply that the maximum-ID model minimizes OOD error; U(b) is minimized at b=Delta and increases beyond it.

full rationale

The theoretical Proposition 1 is a self-contained derivation under the stated half-normal assumption; deriving Cov(T,U)=2k^3(k-alpha*Delta) and the sign condition is not itself circular, and there are no load-bearing self-citations in the manuscript. However, the central empirical claim is circular in an important respect: Section 5.3 explicitly says the Medium/High and highest-ID-error selection rule was chosen 'based on the negative ID-OOD correlation patterns observed in Figure 1', and Figure 1 uses OOD test data. The same OOD test data then appears in Table 1 as the evidence that ADB 'achieves robust mean error reductions'. This is a fitted-input-called-prediction pattern: the selector was tuned to the test-set correlation, then the test-set performance of that selector is presented as confirmation. The paper's own Table 6 further weakens the claim by showing that on MSD with the Cumulative method the Low group has the best OOD performance (PR 95.0, MAE 0.848), directly contradicting the asserted preference for Medium/High groups. The theoretical bridge from ADB's deviation groups to the bias-variation parameter k is also asserted rather than measured, and Section 6 defers the quantitative relationship to future work, but that is an unsupported premise rather than a circular reduction. Overall, because the empirical headline result reduces to re-reporting a test-set pattern used to define the selection rule, the score is 6.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central claim rests on several unverified premises: the half-normal distribution of bias, the mean-shift error model, the bridge between training-sequence deviation and bias variation, and the assumption that label-space W1 distances capture the relevant shift. Free parameters include the ADB thresholds, Sinkhorn regularization, latent dimensionality, and sample sizes, all chosen by hand or quantiles of the experimental data.

free parameters (6)
  • threshold tau_low = 3 (T1S1 Batchwise), 2 (MSD Batchwise), 3 (T1S1 Cumulative), 6 (MSD Cumulative)
    Chosen from quantiles of the empirical outlier-count distribution to split permutations into Low versus Medium groups; varies by dataset and approach (Table 3).
  • threshold tau_high = 6 (T1S1 Batchwise), 7 (MSD Batchwise), 13 (T1S1 Cumulative), 14 (MSD Cumulative)
    Chosen from quantiles of the empirical outlier-count distribution to split permutations into Medium versus High groups (Table 3).
  • Sinkhorn entropy regularization epsilon = 0.05
    Chosen by hand in Section 5.1; affects the debiased distance and therefore all permutation classifications.
  • VAE latent dimension = 32 for T1S1, 8 for MSD
    Chosen architecture; dimensionality reduction is used before distance computation.
  • number of permutations M = 500
    Chosen in Section 5.1; the population of permutations is sampled, affecting the outlier counts and thresholds.
  • number of models per deviation group = 30
    Chosen in Section 5.1; the final model is selected from a sample of 10 from Medium and High groups, so this sample size drives selection variance.
assumptions (4)
  • ad hoc to paper Bias parameter b across candidate models follows a half-normal distribution with scale k.
    Invoked in Section 3.2, Eq. 3 to derive the negative-correlation condition; justified only by mathematical convenience, not by empirical evidence.
  • domain assumption ID error component T(b)=b^2 and OOD error component U(b)=(b-Delta)^2.
    Mean-shift model in Section 3.1; ignores irreducible error, noise, and multi-dimensional bias structure.
  • ad hoc to paper Training-sequence distributional deviation (debiased Sinkhorn distance) induces the theoretical bias-variation parameter k.
    Asserted in Section 4.1 but never measured or validated; no experiment connects the deviation groups to actual k values.
  • domain assumption OOD test sets with label-space W1 distances between 0.61 and 0.86 represent meaningful distribution shifts.
    Section 5.1 and Appendix B.2; assumes label shift is the dominant form of shift and that the W1 distance in label space is a proxy for the theoretical Delta.
invented entities (2)
  • Bias parameter b as a half-normal random variable
    purpose: Theoretical device to define ID and OOD error components and derive the negative-correlation condition.
    Introduced in Section 3.2; never measured or directly estimated from experiments, so no falsifiable handle outside the paper.
  • Low, Medium, and High deviation groups
    purpose: Categorize training permutations by statistical deviation from the global distribution to induce controlled bias diversity.
    Computational constructs with thresholds fitted to the observed outlier counts; their mapping to k is asserted but not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias as a Virtue: Rethinking Generalization under Distribution Shifts." pith.science (2026). https://pith.science/paper/CRAAJNME

@misc{pith2026250600407,
  author       = {Pith},
  title        = {Pith review of: Bias as a Virtue: Rethinking Generalization under Distribution Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CRAAJNME}},
  note         = {Machine review of arXiv:2506.00407}
}
read the original abstract

Machine learning models often degrade when deployed on data distributions different from their training data. Challenging conventional validation paradigms, we demonstrate that higher in-distribution (ID) bias can lead to better out-of-distribution (OOD) generalization. Our Adaptive Distribution Bridge (ADB) framework implements this insight by introducing controlled statistical diversity during training, enabling models to develop bias profiles that effectively generalize across distributions. Empirically, we observe a robust negative correlation where higher ID bias corresponds to lower OOD error--a finding that contradicts standard practices focused on minimizing validation error. Evaluation on multiple datasets shows our approach significantly improves OOD generalization. ADB achieves robust mean error reductions of up to 26.8% compared to traditional cross-validation, and consistently identifies high-performing training strategies, evidenced by percentile ranks often exceeding 74.4%. Our work provides both a practical method for improving generalization and a theoretical framework for reconsidering the role of bias in robust machine learning.

Figures

Figures reproduced from arXiv: 2506.00407 by the authors.

Figure 1
Figure 1. Relationship between in-distribution and out-of-distribution MAE across Low, Medium, [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages

  1. [1]

    Invariant risk mini- mization.arXiv preprint arXiv:1907.02893, 2019

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk mini- mization.arXiv preprint arXiv:1907.02893, 2019. 9

    arXiv 1907

  2. [2]

    A survey of cross-validation procedures for model selection

    Sylvain Arlot and Alain Celisse. A survey of cross-validation procedures for model selection. Statistics surveys, 4:40–79, 2010

    work page 2010

  3. [3]

    Curriculum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. InProceedings of the 26th Annual International Conference on Machine Learning, pages 41–48, 2009

    work page 2009

  4. [4]

    Ellis, Brian Whitman, and Paul Lamere

    Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. The million song dataset. InProceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR), pages 591–596, Miami, Florida, 2011

    work page 2011

  5. [5]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems, volume 26, 2013

    2013

  6. [6]

    Under- specification presents challenges for credibility in modern machine learning.arXiv preprint arXiv:2011.03395, 2020

    Alexander D’Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Under- specification presents challenges for credibility in modern machine learning.arXiv preprint arXiv:2011.03395, 2020

    arXiv 2011

  7. [7]

    A unified bias-variance decomposition.Proceedings of 17th International Conference on Machine Learning, pages 231–238, 2000

    Pedro Domingos. A unified bias-variance decomposition.Proceedings of 17th International Conference on Machine Learning, pages 231–238, 2000

    work page 2000

  8. [8]

    Learning models with uniform performance via distributionally robust optimization

    John C Duchi and Hongseok Namkoong. Learning models with uniform performance via distributionally robust optimization. InAdvances in Neural Information Processing Systems, volume 31, 2018

    work page 2018

Show all 39 references
  1. [9]

    Domain-adversarial training of neural networks.The Journal of Machine Learning Research, 17(1):2096–2030, 2016

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks.The Journal of Machine Learning Research, 17(1):2096–2030, 2016

    2016

  2. [10]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. InInternational Conference on Machine Learning, pages 1263–1272. PMLR, 2017

    2017

  3. [11]

    Au- tomated curriculum learning for neural networks

    Alex Graves, Marc G Bellemare, Jacob Menick, Remi Munos, and Koray Kavukcuoglu. Au- tomated curriculum learning for neural networks. InInternational Conference on Machine Learning, pages 1311–1320. PMLR, 2017

    2017

  4. [12]

    In search of lost domain generalization

    Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. InInternational Conference on Learning Representations, 2021

    2021

  5. [13]

    On the power of curriculum learning in training deep networks

    Guy Hacohen and Daphna Weinshall. On the power of curriculum learning in training deep networks. InInternational Conference on Machine Learning, pages 2535–2544. PMLR, 2019

    2019

  6. [14]

    Feature-level ensembling for both accuracy and robustness

    Pavel Izmailov, Dmitry Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Feature-level ensembling for both accuracy and robustness. InInternational Conference on Machine Learning, pages 9981–9991. PMLR, 2022

    2022

  7. [15]

    Variance and bias for general loss functions.Machine learning, 51(2):115–135, 2003

    Gareth M James. Variance and bias for general loss functions.Machine learning, 51(2):115–135, 2003

    2003

  8. [16]

    Grad-match: Gradient matching based data subset selection for efficient deep model training

    Krishnateja Killamsetty, Durga Sivasubramanian, Ganesh Ramakrishnan, and Rishabh Iyer. Grad-match: Gradient matching based data subset selection for efficient deep model training. InInternational Conference on Machine Learning, pages 5464–5474. PMLR, 2021

    2021

  9. [17]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations, 2014

    2014

  10. [18]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. InInternational Conference on Learning Representations, 2013

    2013

  11. [19]

    Wilds: A benchmark of in-the-wild distribution shifts.International Conference on Machine Learning, pages 5637–5664, 2021

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Xu, Percy Liang, et al. Wilds: A benchmark of in-the-wild distribution shifts.International Conference on Machine Learning, pages 5637–5664, 2021. 10

    2021

  12. [20]

    Fine-tuning can distort pretrained features and underperform out-of-distribution

    Ananya Kumar, Aditi Raghunathan, Roy Jones, Tengyu Ma, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. InInternational Conference on Learning Representations, 2022

    2022

  13. [21]

    Rdkit: Open-source cheminformatics

    Greg Landrum et al. Rdkit: Open-source cheminformatics. http://www.rdkit.org,

  14. [22]

    Improved complexity analysis of the sinkhorn and greenkhorn algorithms for optimal transport.arXiv preprint arXiv:2305.14939, 2023

    Jianzhou Luo, Dingchuan Yang, and Ke Wei. Improved complexity analysis of the sinkhorn and greenkhorn algorithms for optimal transport.arXiv preprint arXiv:2305.14939, 2023

    2023 arXiv

  15. [23]

    Prioritized training on points that are learnable, worth learning, and not yet learnt

    Sören Mindermann, Jan Markus Brauner, Muhammad Tarek Razzak, Mrinank Sharma, Andreas Kirsch, Winnie Xu, Benedikt Höltgen, Aidan N Gomez, Adrien Morisot, Sebastian Farquhar, and Yarin Gal. Prioritized training on points that are learnable, worth learning, and not yet learnt. In...

    2022

  16. [24]

    A unifying view on dataset shift in classification.Pattern recognition, 45(1):521–530, 2012

    Jose G Moreno-Torres, Troy Raeder, Rocío Alaiz-Rodríguez, Nitesh V Chawla, and Francisco Herrera. A unifying view on dataset shift in classification.Pattern recognition, 45(1):521–530, 2012

    2012

  17. [25]

    The generation of a unique machine description for chemical structures- a technique developed at chemical abstracts service.Journal of Chemical Documentation, 5(2):107–113, 1965

    Harry L Morgan. The generation of a unique machine description for chemical structures- a technique developed at chemical abstracts service.Journal of Chemical Documentation, 5(2):107–113, 1965

    1965

  18. [26]

    Jonas Peters, Peter Bühlmann, and Nicolai Meinshausen. Causal inference using invariant prediction: identification and confidence intervals.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 78(5):947–1012, 2016

    2016

  19. [27]

    Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. InInternational Conference on Learning Representations, 2019

    2019

  20. [28]

    An investigation of why overparameterization exacerbates spurious correlations

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. InInternational Conference on Machine Learning, pages 8346–8356. PMLR, 2020

    2020

  21. [29]

    Breeds: Benchmarks for sub- population shift

    Shibani Santurkar, Dimitris Tsipras, and Aleksander Madry. Breeds: Benchmarks for sub- population shift. InAdvances in Neural Information Processing Systems, volume 34, pages 10816–10830, 2021

    2021

  22. [30]

    Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    2021

  23. [31]

    Active learning for convolutional neural networks: A core-set approach

    Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. InInternational Conference on Learning Representations, 2018

    2018

  24. [32]

    Active learning literature survey

    Burr Settles. Active learning literature survey. Technical report, University of Wisconsin- Madison Department of Computer Sciences, 2009

    2009

  25. [33]

    Towards out-of-distribution generalization: A survey.arXiv preprint arXiv:2108.13624, 2021

    Zheyan Shen, Jiashuo Liu, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey.arXiv preprint arXiv:2108.13624, 2021

    2021 arXiv

  26. [34]

    Measuring robustness to natural distribution shifts in image classification

    Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. InAdvances in Neural Information Processing Systems, volume 33, pages 18583–18599, 2020

    2020

  27. [35]

    Evading the simplicity bias: Training robust ensembles with random weights

    Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Evading the simplicity bias: Training robust ensembles with random weights. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15410–15419, 2022

    2022

  28. [36]

    Adversarial discriminative domain adaptation

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7167–7176, 2017. 11

    2017

  29. [37]

    A fine-grained analysis on distribution shift

    Olivia Wiles, Sven Gowal, Florian Stimberg, Sylvestre-Alvise Rebuffi, Ira Ktena, Krishnamurthy Dvijotham, and Ali Taylan Cemgil. A fine-grained analysis on distribution shift. InInternational Conference on Machine Learning, pages 23692–23710. PMLR, 2022

  30. [38]

    Data-centric artificial intelligence: A survey.IEEE Transactions on Knowledge and Data Engineering, 35(12):11442–11458, 2023

    Daochen Zha, Zaid Pervaiz Bhat, Kwei-Herng Lai, Fan Yang, Zhimeng Jiang, Shaochen Zhong, and Xia Hu. Data-centric artificial intelligence: A survey.IEEE Transactions on Knowledge and Data Engineering, 35(12):11442–11458, 2023. A Mathematical Analysis of Negative Correlation Th...

    2023

  31. [2013]

    Accessed: 2023-12-01

    2023

Pith tools

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