Pith. sign in

REVIEW 3 major objections 6 minor 62 references

Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Removing linear correlations between target and bias features before the final linear classifier significantly mitigates bias and matches or beats prior methods on four benchmarks, with a single default coefficient.

desk verdict Strong empirical debiasing method whose stated mechanism is not actually established; worth engaging, needs revision. read the letter →

arxiv 2507.20284 v1 pith:3GBEHKEV submitted 2025-07-27 cs.CV cs.LG

Yooshin Cho , Hanbyel Cho , Janghyeon Lee , HyeongGwon Hong , Jaesung Ahn , Junmo Kim This is my paper
classification cs.CVcs.LG
keywords featurewhiteningbiasmitigationspuriouscorrelationsdemographicparityequalizedoddslineardecorrelationlast-layerretraininggrouprobustness
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

The paper claims that a large share of what makes biased classifiers biased is linear correlation between the features that feed the last linear layer, and that deliberately whitening those features is enough to reduce the bias in practice. The authors propose Controllable Feature Whitening (CFW), which concatenates target and bias features, whitens them using a covariance matrix that blends biased and re-weighted unbiased statistics, and then trains separate linear classifiers on the whitened target and whitened bias channels. This replaces adversarial learning and regularization terms, which are often unstable and hyperparameter-sensitive. Across Corrupted CIFAR-10, Biased FFHQ, WaterBirds, and Celeb-A, the method is claimed to achieve state-of-the-art or competitive performance while a single coefficient of 0.25 works consistently.

What carries the argument

The central mechanism is the controllable whitening module $W_{\lambda}$, defined by $z_w = \Sigma_{\lambda}^{-1/2}(z - \mu)$, applied to $z = [z_t; z_b]$. Here $\Sigma_b$ is the covariance of the concatenated features on the biased training set, $\Sigma_u$ is the same covariance re-weighted so all $(Y, B)$ groups have equal probability, and $\Sigma_{\lambda} = \lambda \Sigma_u + (1-\lambda) \Sigma_b$. The inverse square root is computed with coupled Newton-Schultz iterations, which the paper uses both for numerical stability and for the freedom to choose a unitary rotation so that $z_{wt}$ stays close to $z_t$. This one module does the work of an adversarial discriminator: by making the whitened target and bias channels linearly independent, it prevents a linear layer from predicting one from the other, and the $\lambda$ knob interpolates between demographic parity and equalized odds.

What would settle it

Train CFW on Biased FFHQ with $\lambda = 0$, freeze the module, and fit a fresh linear probe to predict gender from the whitened target features $z_{wt}$ alone. If this probe reaches substantially above-chance accuracy while $z_{wt}$ and $z_{wb}$ remain orthogonal by construction, the claimed mechanism would be refuted, since the bias would remain linearly readable in the features actually fed to the target classifier.

Watch

Extended reading notes

Core claim

The central discovery is that satisfying linear independence between the two halves of a whitened feature vector is sufficient to stop a last-layer linear classifier from exploiting a known bias attribute, even though linear independence is weaker than statistical independence. The paper demonstrates this by extracting a target feature $z_t$ from a frozen pretrained encoder and a bias feature $z_b$ from a separately trained bias encoder, then whitening the concatenated vector. After whitening with the coupled Newton-Schultz iteration, the whitened target features $z_{wt}$ and whitened bias features $z_{wb}$ have zero covariance, so neither can be predicted by a linear map from the other. Training a bias classifier on $z_{wb}$ forces $z_{wt}$ to forget bias information in the linear regime, and training the target classifier on $z_{wt}$ yields predictions that barely track the bias attribute. The paper also claims that re-weighting the covariance matrix over the unbiased group distribution turns the same mechanism into an equalized-odds regularizer, and that interpolating biased and unbiased covariances with a coefficient $\lambda \in [0,1]$ trades off demographic parity against task utility, with $\lambda = 0.25$ a reliable default.

Load-bearing premise

The make-or-break premise is that zeroing out the linear correlation between the whitened target features and whitened bias features is enough to stop the final classifier from using the bias attribute, even though this is weaker than full statistical independence and the whitening transform mixes both original feature sets.

Editorial extensions

If this is right

  • If linear decorrelation is sufficient, bias mitigation reduces to a fixed whitening step applied only to the features before the last linear layer, leaving the pretrained target encoder frozen.
  • The same whitening module, with covariance re-weighting, addresses both demographic parity and equalized odds: small $\lambda$ enforces demographic parity, large $\lambda$ enforces equalized odds, and intermediate values give a smooth trade-off.
  • Because no adversarial network or regularization penalty is needed, training is stable and uses a single default coefficient ($\lambda = 0.25$) that works across all four benchmark datasets.
  • Combining the whitening module with a better pretrained representation, such as a SelecMix-pretrained target encoder, further improves results, so the method composes with representation-quality improvements.
  • The method improves worst-group accuracy and shrinks the gap between unbiased and bias-conflicting accuracy on Celeb-A and WaterBirds, and it outperforms last-layer retraining approaches when the backbone is not ImageNet-pretrained.

Reading between the lines

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

  • A testable extension is to apply the same whitening operation to text or tabular tasks whose final classifier is also linear, since the claimed mechanism depends only on linear readability of the bias in the final features.
  • The paper does not test this, but the mechanism implies that the bias removed is only what the bias encoder $h_b$ can linearize; a weaker bias encoder could leave bias in $z_{wt}$ even when $z_{wt}$ and $z_{wb}$ are orthogonal.
  • The claim that $\lambda = 0.25$ is hyperparameter-free is empirical across four datasets; a stronger test would sweep $\lambda$ on additional group-shift benchmarks to see whether the default holds beyond the presented suite.
  • Because whitening acts on the covariance of the current batch, the method could be sensitive to batch composition; a natural extension would be to test whether the same $\lambda$ remains effective under severe group imbalance or small batch sizes.
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 Controllable Feature Whitening (CFW), a debiasing method that whitens the concatenation of a frozen target-encoder feature z_t and a trained bias-encoder feature z_b, then trains separate linear classifiers on the two whitened blocks z_wt and z_wb to predict the target and bias attributes, respectively. The covariance matrix used for whitening is a weighted average of the biased and reweighted (class-balanced) covariance estimates, with weight λ, so that λ=0 targets demographic parity and λ=1 targets equalized odds. The authors set λ=0.25 on bFFHQ and use it across all datasets, claiming the method is hyperparameter-free. They report state-of-the-art or competitive results on Corrupted CIFAR-10, Biased FFHQ, WaterBirds, and Celeb-A, including worst-group accuracy and equalized-odds metrics, and provide ablations on loss weighting and on the choice of inverse-square-root solver.

Significance. The empirical contribution is substantial: CFW obtains the best or second-best worst-group accuracy on WaterBirds, strong bias-conflicting accuracy on bFFHQ and Celeb-A, and consistently low equalized-odds violation on Celeb-A, with multiple random seeds and ablations. The method is simple, avoids adversarial training, and the use of a single λ=0.25 across datasets is a meaningful practical convenience. If the linear-decorrelation mechanism is accepted, the paper offers a useful addition to the debiasing toolkit. However, the paper's central theoretical justification, that linear decorrelation between whitened target and bias features suffices to remove linear bias information from the target branch, is not established; this weakens the claimed connection between the mechanism and the strong empirical results. The paper does not mention code release, which would further aid reproducibility.

major comments (3)
  1. [Sec. 4.1 (after Eq. 6)] The claim that, because z_wt and z_wb are linearly independent and g_wb is trained to predict B from z_wb, the whitened target feature z_wt "becomes incapable of linearly encoding bias information" does not follow from the construction. Whitening enforces Cov(z_wt, z_wb)=0, not Cov(z_wt, B)=0. Since the whitening map is invertible, the best linear predictor of B in the full whitened space generally has nonzero coefficients on both blocks; nothing in Eq. (6) penalizes the linear predictability of B from z_wt. This is not merely a formal gap: in a simple one-dimensional example with jointly Gaussian z_t and z_b, the whitened target block retains nonzero correlation with B unless B is exactly a linear function of z_wb. Please provide direct evidence for the mechanism, e.g., train a linear probe on z_wt to predict B and report its accuracy against chance before and after whitening, or give a formal condition under which the claim holds. The current indirect evidence (t-SNE in Fig. 5 and group accuracies) does not discriminate between genuine removal of bias information from z_wt and other explanations such as the reweighted covariance reducing the bias-aligned signal.
  2. [Secs. 4.2, 5.3, and Table 4] The statement in Sec. 4.2 that "demographic parity and equalized odds become equivalent in an unbiased distribution" is true as a property of the data distribution, but it does not imply that whitening with Σ_u "naturally promotes" equalized odds. Equalized odds is a property of the predictor, not of the whitening covariance alone, and the classifier is still trained on the biased dataset. The empirical ΔEO results in Fig. 3 are encouraging, but the explanation should be reframed as an empirical observation or supported by an explicit derivation.
  3. [Secs. 4.1, 5.3, and Table 4] The "hyperparameter-free" claim is not fully supported because the loss-weighting (LW) scheme is a separate component used in the main results but never specified. Table 4 shows that the full method includes "+ LW", and Sec. 5.1 says LW under-weights bias-aligned samples, but the paper does not state how the per-group weights are computed or whether any strength parameter is involved. If LW has a tunable weight, the title and abstract overstate the hyperparameter-free property; if it is fixed, please provide the exact recipe in the appendix. At minimum, the abstract should be qualified to "no per-dataset tuning of λ" and the LW details should be supplied for reproducibility.
minor comments (6)
  1. [Abstract] The word "systemically" should be "systematically".
  2. [Sec. 5.1 and Fig. 2 caption] Typo "stabilizeining" should be "stabilizing"; also use "bias-conflicting" consistently in the caption text.
  3. [Table 6 and Abstract] The abstract's "outperforms existing approaches on four benchmark datasets" is too strong because on WaterBirds the mean accuracy of Ours (96.82) is below Vanilla (98.1), as the paper itself notes; please restrict the claim to worst-group or bias-conflicting performance or qualify the mean-accuracy comparison.
  4. [Appendix B] The paper does not specify whether the whitening statistics at test time are batch statistics or running averages; since whitening is data-dependent, this should be stated for reproducibility.
  5. [Table 3] The shorthand "T=a/S=m" and similar entries in Table 3 are not defined in the caption; please clarify the naming convention.
  6. [References] References [38] and [39] are duplicates of the same Celeb-A paper; one should be removed and the citations merged.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the whitening construction is fixed and externally evaluated; the only mild tuning is the single λ value selected on bFFHQ.

full rationale

The central method is not derived from the benchmark results. CFW is a fixed transformation: concatenated target/bias features are whitened with Σλ = λΣu + (1−λ)Σb, split, and followed by linear classifiers trained with cross-entropy (Eq. 6). The reported metrics (unbiased, bias-conflicting, worst-group accuracy, ∆DP, ∆EO) are evaluated on held-out test data, so the evaluation is not an input to the construction. The single coefficient λ = 0.25 is chosen empirically on bFFHQ and then transferred to other datasets; this is hyperparameter selection, not a fitted quantity that is renamed as a prediction, and it does not make the reported gains forced by construction. The paper's load-bearing claim that zero covariance between zwt and zwb makes zwt incapable of linearly encoding B is an empirical/analytic hypothesis: whitening enforces Cov(zwt, zwb)=0, not Cov(zwt, B)=0, and the paper does not prove the implication. That is a potential correctness gap, but it is not circularity, because the construction does not assume the conclusion. Self-citations (refs. 10, 29, 32) are used for technical background or backbone training, not as the only support for the central claim, and no uniqueness theorem is imported from the authors' prior work. The derivation chain therefore does not reduce to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on heuristics rather than derived bounds. The free parameters are few and mostly fixed a priori, but the core assumption that linear decorrelation is sufficient for bias mitigation is unproven. The hyperparameter-free claim obscures the selection of lambda on bFFHQ.

free parameters (3)
  • lambda (weight coefficient) = 0.25
    Selected based on bFFHQ experiments (Section 5.1) and then fixed for all other datasets. The paper calls the method hyperparameter-free despite this tuning.
  • T (Newton-Schultz iterations) = 5
    Taken from prior work [20,55], not fitted in this paper, but still a training hyperparameter that affects the whitening approximation.
  • Loss weighting (LW) scheme = unspecified
    The paper says LW under-weights bias-aligned samples to mimic the unbiased loss (Section 5.1), but no exact formula or weighting values are provided, leaving an unstated choice for re-implementation.
assumptions (4)
  • domain assumption Linear independence between the whitened feature blocks is sufficient to prevent bias transfer through the last linear layer.
    Central premise of the method, stated in Section 4.1. No proof is given; the paper relies on empirical validation.
  • domain assumption The unbiased distribution can be approximated by a uniform joint distribution over Y and B, P(y,b|D_u)=1/(N_Y*N_B).
    Modeling choice in Eq 8. The true unbiased marginals are unknown, and re-weighting group covariances only approximates the unbiased covariance.
  • domain assumption Freezing the target encoder and only training the bias encoder and linear heads is enough to achieve fairness.
    Motivated by prior last-layer retraining works [33,52], but not guaranteed to interact correctly with the changing whitening transform.
  • standard math Batch-level covariance estimates provide reliable whitening statistics.
    Standard practice in batch normalization, but the paper does not specify test-time handling (running statistics versus recomputed batch statistics).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation." pith.science (2026). https://pith.science/paper/3GBEHKEV

@misc{pith2026250720284,
  author       = {Pith},
  title        = {Pith review of: Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GBEHKEV}},
  note         = {Machine review of arXiv:2507.20284}
}
read the original abstract

As the use of artificial intelligence rapidly increases, the development of trustworthy artificial intelligence has become important. However, recent studies have shown that deep neural networks are susceptible to learn spurious correlations present in datasets. To improve the reliability, we propose a simple yet effective framework called controllable feature whitening. We quantify the linear correlation between the target and bias features by the covariance matrix, and eliminate it through the whitening module. Our results systemically demonstrate that removing the linear correlations between features fed into the last linear classifier significantly mitigates the bias, while avoiding the need to model intractable higher-order dependencies. A particular advantage of the proposed method is that it does not require regularization terms or adversarial learning, which often leads to unstable optimization in practice. Furthermore, we show that two fairness criteria, demographic parity and equalized odds, can be effectively handled by whitening with the re-weighted covariance matrix. Consequently, our method controls the trade-off between the utility and fairness of algorithms by adjusting the weighting coefficient. Finally, we validate that our method outperforms existing approaches on four benchmark datasets: Corrupted CIFAR-10, Biased FFHQ, WaterBirds, and Celeb-A.

Figures

Figures reproduced from arXiv: 2507.20284 by the authors.

Figure 1
Figure 1. Overview of proposed method. X is the mini-batch of the images that sampled from the biased training dataset. ht(·) and hb(·) are target and bias encoder, respectively. To reduce the dependency, we remove the linear correlation between the target feature zt and bias feature zb using the controllable whitening module Wλ(·), that can handle demographic parity and equalized odds by controlling the coefficient λ. Subseq… view at source ↗
Figure 2
Figure 2. The solid lines and dashed lines in (a) are the test accu [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of ∆DP and ∆EO with respect to training iterations. It shows that whitening with the unbiased covariance (λ=1) successfully regularizes ∆EO. By contrast, whitening with the biased covariance (λ=0) successfully regularizes the ∆DP . 5 15 25 35 45 Iters (# 103 ) 0.25 0.5 0.75 Loss 6 = 0.0 6 = 0.25 6 = 0.5 6 = 0.75 6 = 1.0 (a) Training Loss (bias-aligned) 0 0.25 0.5 0.75 1 6 60 70 80 90 100 Test Accuracy (… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Illustration of the 2D projected target features [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 49 canonical work pages

  1. [1]

    A closer look at memorization in deep networks

    Devansh Arpit, Stanisław Jastrz˛ ebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In International conference on machine learning , pages 233–

    work page

  2. [2]

    Learning de-biased representations with biased representations

    Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International Conference on Ma- chine Learning, pages 528–539. PMLR, 2020. 1, 7

    work page 2020

  3. [3]

    Mitigating biases in blackbox feature extractors for image classification tasks

    Abhipsa Basu, Saswat Subhajyoti Mallick, et al. Mitigating biases in blackbox feature extractors for image classification tasks. Advances in Neural Information Processing Systems, 37:106411–106439, 2025. 7

    work page 2025

  4. [4]

    indepen- dent components

    Anthony J Bell and Terrence J Sejnowski. The “indepen- dent components” of natural scenes are edge filters. Vision research, 37(23):3327–3338, 1997. 3, 8

    work page 1997

  5. [5]

    Al- gorithms for the matrix pth root

    Dario A Bini, Nicholas J Higham, and Beatrice Meini. Al- gorithms for the matrix pth root. Numerical Algorithms, 39 (4):349–378, 2005. 3

    work page 2005

  6. [6]

    Rubi: Reducing unimodal biases for visual question answering

    Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. Rubi: Reducing unimodal biases for visual question answering. Advances in neural information pro- cessing systems, 32, 2019. 3

    work page 2019

  7. [7]

    Building classifiers with independency constraints

    Toon Calders, Faisal Kamiran, and Mykola Pechenizkiy. Building classifiers with independency constraints. In 2009 IEEE International Conference on Data Mining Workshops, pages 13–18. IEEE, 2009. 3

    work page 2009

  8. [8]

    Fast model debias with machine un- learning

    Ruizhe Chen, Jianfei Yang, Huimin Xiong, Jianhong Bai, Tianxiang Hu, Jin Hao, Yang Feng, Joey Tianyi Zhou, Jian Wu, and Zuozhu Liu. Fast model debias with machine un- learning. Advances in Neural Information Processing Sys- tems, 36:14516–14539, 2023. 7

    work page 2023

Show all 62 references
  1. [9]

    Better may not be fairer: A study on subgroup discrepancy in image classification

    Ming-Chang Chiu, Pin-Yu Chen, and Xuezhe Ma. Better may not be fairer: A study on subgroup discrepancy in image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4956–4966, 2023. 3

    2023

  2. [10]

    Improving generalization of batch whitening by convo- lutional unit optimization

    Yooshin Cho, Hanbyel Cho, Youngsoo Kim, and Junmo Kim. Improving generalization of batch whitening by convo- lutional unit optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5321– 5329, 2021. 3

    2021

  3. [11]

    Flexibly fair representation learning by disentan- glement

    Elliot Creager, David Madras, Jörn-Henrik Jacobsen, Marissa Weis, Kevin Swersky, Toniann Pitassi, and Richard Zemel. Flexibly fair representation learning by disentan- glement. In International conference on machine learning , pages 1436–1445. PMLR, 2019. 1

    2019

  4. [12]

    Cholesky factor- ization of matrices in parallel and ranking of graphs

    Dariusz Dereniowski and Marek Kubale. Cholesky factor- ization of matrices in parallel and ranking of graphs. In In- ternational conference on parallel processing and applied mathematics, pages 985–992. Springer, 2003. 3, 8

    2003

  5. [13]

    Whitening for self-supervised representation learning

    Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In International Conference on Machine Learning, pages 3015–3024. PMLR, 2021. 3

    2021

  6. [14]

    On formalizing fair- ness in prediction with machine learning

    Pratik Gajane and Mykola Pechenizkiy. On formalizing fair- ness in prediction with machine learning. arXiv preprint arXiv:1710.03184, 2017. 4

    2017 arXiv

  7. [15]

    Equality of op- portunity in supervised learning

    Moritz Hardt, Eric Price, and Nati Srebro. Equality of op- portunity in supervised learning. Advances in neural infor- mation processing systems, 29, 2016. 2, 4

    2016

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 1

    2016

  9. [17]

    Newton’s method for the matrix square root

    Nicholas J Higham. Newton’s method for the matrix square root. Mathematics of computation, 46(174):537–549, 1986. 3, 8

    1986

  10. [18]

    Functions of matrices: theory and com- putation

    Nicholas J Higham. Functions of matrices: theory and com- putation. SIAM, 2008. 3, 8

    2008

  11. [19]

    Decorre- lated batch normalization

    Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorre- lated batch normalization. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 791–800, 2018. 1

    2018

  12. [20]

    It- erative normalization: Beyond standardization towards ef- ficient whitening

    Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. It- erative normalization: Beyond standardization towards ef- ficient whitening. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4874–4883, 2019. 1, 8

    2019

  13. [21]

    An investigation into the stochasticity of batch whiten- ing

    Lei Huang, Lei Zhao, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. An investigation into the stochasticity of batch whiten- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 6439–6448,

  14. [22]

    Selecmix: Debiased learning by mixing up contradicting pairs

    Inwoo Hwang, Sangjun Lee, Yunhyeok Kwak, Seong Joon Oh, Damien Teney, Jin-Hwa Kim, and Byoung-Tak Zhang. Selecmix: Debiased learning by mixing up contradicting pairs. In ICML 2022: Workshop on Spurious Correlations, Invariance and Stability, 2022. 7

    2022

  15. [23]

    Neu- ral tangent kernel: Convergence and generalization in neural networks

    Arthur Jacot, Franck Gabriel, and Clément Hongler. Neu- ral tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing sys- tems, 31, 2018. 3

    2018

  16. [24]

    Fair feature distillation for visual recognition

    Sangwon Jung, Donggyu Lee, Taeeon Park, and Taesup Moon. Fair feature distillation for visual recognition. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 12115–12124, 2021. 8

    2021

  17. [25]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4401–4410, 2019. 2, 12

    2019

  18. [26]

    Optimal whitening and decorrelation

    Agnan Kessy, Alex Lewin, and Korbinian Strimmer. Optimal whitening and decorrelation. The American Statistician, 72 (4):309–314, 2018. 3, 8

    2018

  19. [27]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661–18673,

  20. [28]

    Avoiding discrimination through causal reason- ing

    Niki Kilbertus, Mateo Rojas Carulla, Giambattista Paras- candolo, Moritz Hardt, Dominik Janzing, and Bernhard 9 Schölkopf. Avoiding discrimination through causal reason- ing. Advances in neural information processing systems, 30,

  21. [29]

    Learning not to learn: Training deep neural networks with biased data

    Byungju Kim, Hyunwoo Kim, Kyungsu Kim, Sungjin Kim, and Junmo Kim. Learning not to learn: Training deep neural networks with biased data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9012–9020, 2019. 1, 3, 8

    2019

  22. [30]

    Biaswap: Removing dataset bias with bias-tailored swapping augmen- tation

    Eungyeup Kim, Jihyeon Lee, and Jaegul Choo. Biaswap: Removing dataset bias with bias-tailored swapping augmen- tation. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 14992–15001, 2021. 2, 3, 5, 12

    2021

  23. [31]

    Learning debiased classifier with biased committee

    Nayeong Kim, Sehyun Hwang, Sungsoo Ahn, Jaesik Park, and Suha Kwak. Learning debiased classifier with biased committee. Advances in Neural Information Processing Sys- tems, 35:18403–18415, 2022. 3, 5, 7, 12

    2022

  24. [32]

    Nlnl: Negative learning for noisy labels

    Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. Nlnl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 101–110, 2019. 1

    2019

  25. [33]

    Last layer re-training is sufficient for robustness to spu- rious correlations

    Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wil- son. Last layer re-training is sufficient for robustness to spu- rious correlations. arXiv preprint arXiv:2204.02937, 2022. 4, 7, 13

    2022 arXiv

  26. [34]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, Uni- versity of Toronto, Toronto, Ontario, 2009. 2, 5, 12

    2009

  27. [35]

    Deep neural networks as gaussian processes

    Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. arXiv preprint arXiv:1711.00165, 2017. 3

    2017 arXiv

  28. [36]

    Learning debiased representation via disentan- gled feature augmentation

    Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentan- gled feature augmentation. Advances in Neural Information Processing Systems, 34:25123–25133, 2021. 1, 3, 7, 12

    2021

  29. [37]

    Just train twice: Improving group robustness without training group information

    Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghu- nathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Confer- ence on Machine Learning, pages 6781–6792. PMLR,...

    2021

  30. [38]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730–3738, 2015. 2, 12

    2015

  31. [39]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5

    2015

  32. [40]

    Learning from failure: De-biasing classifier from biased classifier

    Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Pro- cessing Systems, 33:20673–20684, 2020. 1, 2, 3, 5, 7, 12, 13

    2020

  33. [41]

    Spread spurious attribute: Improving worst-group accu- racy with spurious attribute estimation

    Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accu- racy with spurious attribute estimation. arXiv preprint arXiv:2204.02070, 2022. 7, 13

    2022 arXiv

  34. [42]

    Learning disentangled representation for fair facial attribute classification via fairness-aware information align- ment

    Sungho Park, Sunhee Hwang, Dohyung Kim, and Hyeran Byun. Learning disentangled representation for fair facial attribute classification via fairness-aware information align- ment. In Proceedings of the AAAI conference on artificial intelligence, pages 2403–2411, 2021. 8

    2021

  35. [43]

    Fair contrastive learn- ing for facial attribute classification

    Sungho Park, Jewook Lee, Pilhyeon Lee, Sunhee Hwang, Dohyung Kim, and Hyeran Byun. Fair contrastive learn- ing for facial attribute classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10389–10398, 2022. 8

    2022

  36. [44]

    Gradient reversal against discrimination: A fair neural network learning approach

    Edward Raff and Jared Sylvester. Gradient reversal against discrimination: A fair neural network learning approach. In 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) , pages 189–198. IEEE,

    2018

  37. [45]

    Learning unbiased representations via mutual information backpropagation

    Ruggero Ragonesi, Riccardo V olpi, Jacopo Cavazza, and Vittorio Murino. Learning unbiased representations via mutual information backpropagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2729–2738, 2021. 1

    2021

  38. [46]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD interna- tional conference on knowledge discovery and data mining , pages 1135–1144, 2016. 1

    2016

  39. [47]

    Fairbatch: Batch selection for model fairness

    Yuji Roh, Kangwook Lee, Steven Euijong Whang, and Changho Suh. Fairbatch: Batch selection for model fairness. arXiv preprint arXiv:2012.01696, 2020. 3

    2012 arXiv

  40. [48]

    Unsuper- vised domain adaptation using feature-whitening and con- sensus loss

    Subhankar Roy, Aliaksandr Siarohin, Enver Sangineto, Samuel Rota Bulo, Nicu Sebe, and Elisa Ricci. Unsuper- vised domain adaptation using feature-whitening and con- sensus loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9471– ...

    2019

  41. [49]

    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. arXiv preprint arXiv:1911.08731, 2019. 2, 5, 7, 12, 13

    1911 arXiv

  42. [50]

    Information-theoretic bias reduction via causal view of spu- rious correlation

    Seonguk Seo, Joon-Young Lee, and Bohyung Han. Information-theoretic bias reduction via causal view of spu- rious correlation. arXiv preprint arXiv:2201.03121 , 2022. 2

    2022 arXiv

  43. [51]

    Whitening and coloring batch transform for gans

    Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening and coloring batch transform for gans. arXiv preprint arXiv:1806.00420, 2018. 3

    2018 arXiv

  44. [52]

    End: Entangling and disentangling deep rep- resentations for bias correction

    Enzo Tartaglione, Carlo Alberto Barbano, and Marco Grangetto. End: Entangling and disentangling deep rep- resentations for bias correction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13508–13517, 2021. 4, 5, 7, 12

    2021

  45. [53]

    Learning robust representations by projecting super- ficial statistics out

    Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting super- ficial statistics out. arXiv preprint arXiv:1903.06256, 2019. 7 10

    1903 arXiv

  46. [54]

    Improving out-of-distribution robustness via selective augmentation

    Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. Improving out-of-distribution robustness via selective augmentation. In International Con- ference on Machine Learning, pages 25407–25437. PMLR,

  47. [55]

    Network deconvolution

    Chengxi Ye, Matthew Evanusa, Hua He, Anton Mitrokhin, Tom Goldstein, James A Yorke, Cornelia Fermüller, and Yiannis Aloimonos. Network deconvolution. arXiv preprint arXiv:1905.11926, 2019. 3, 8

    1905 arXiv

  48. [56]

    Wide residual net- works

    Sergey Zagoruyko and Nikos Komodakis. Wide residual net- works. arXiv preprint arXiv:1605.07146, 2016. 1

    2016 arXiv

  49. [57]

    Mitigating unwanted biases with adversarial learning

    Brian Hu Zhang, Blake Lemoine, and Margaret Mitchell. Mitigating unwanted biases with adversarial learning. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 335–340, 2018. 4

    2018

  50. [58]

    Contrastive adapters for foundation model group robustness

    Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. Advances in Neural In- formation Processing Systems, 35:21682–21697, 2022. 7

    2022

  51. [59]

    Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations

    Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. 3, 7, 13

    2022 arXiv

  52. [60]

    Generalized cross entropy loss for training deep neural networks with noisy labels

    Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Ad- vances in neural information processing systems , 31, 2018. 3

    2018

  53. [61]

    Learning bias-invariant representation by cross-sample mutual information minimization

    Wei Zhu, Haitian Zheng, Haofu Liao, Weijian Li, and Jiebo Luo. Learning bias-invariant representation by cross-sample mutual information minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 15002–15012, 2021. 1, 3, 7, 12

    2021

  54. [62]

    Ob- ject recognition with and without objects

    Zhuotun Zhu, Lingxi Xie, and Alan L Yuille. Ob- ject recognition with and without objects. arXiv preprint arXiv:1611.06596, 2016. 1 11 Supplementary Material for Controllable Feature Whitening for Hyperparameter-Free Bias Mitigation A. Datasets Corrupted CIFAR-10 was proposed ...

    2016 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.