acceptodds
Under review as a conference paper at ICLR 2027

Spectral Suppression for Annotation-Free Bias Mitigation

Abstract

Deep neural networks trained with Empirical Risk Minimization (ERM) often rely on spurious correlations rather than robust features, particularly when mul- tiple bias attributes are strongly correlated with the target label. Existing debias- ing methods typically require expensive per-sample bias annotations, which limits their practicality in real-world settings. We introduce Spectral Alignment Sup- pression (SAS), a novel annotation-free debiasing method. SAS identifies the dominant spectral subspace of the centered feature matrix and attenuates classifier feature coordinates with high leverage in that subspace. This is combined with a cosine classifier that makes the logits invariant to uniform feature scaling and per- class prototype scaling. Across Biased-MNIST, CelebA, and MS-COCO, SAS improves substantially over ERM, surpasses annotation-free two-step baselines (JTT, LfF) in the majority of evaluated settings, and remains competitive with su- pervised debiasing methods without using bias annotations. A balanced spectral audit shows dataset-dependent geometry: bias dominates the leading subspace on Biased-MNIST, co-locates with target information on CelebA, and is strongly de- codable together with the target in a compact top-32 subspace on MS-COCO. On unbiased benchmarks (CIFAR-10, STL-10, and MNIST), SAS maintains accuracy comparable to ERM.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Bias Mitigation, Spurious Correlations, Representation Learning, Spectral Analysis, Robust Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2fHpOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper provides a new bias-label free method to train models under spurious correlations, outperforming ERM and several anootation-free two-stage methods and also methods that utilize annotations. The propsoed method, SAS, focuses on subspace learned by the feature encoder, and updates/regularizes the classifier weights after each gradient step and outputs a debiased classifier.

Critical strengths
  • This paper proposes a cheap alternative to methods on mitigating spurious correlations that rely on bias labels that are hard to obtain in practice, and two-stage methods that require twice the training time as ERM methods. The propose algorithm requires one extra step after each gradient step, which applies SVD on the centered batch features with small additional cost. The paper is also well written and very easy to follow.
Critical weaknesses
  • Various works separate the feature encoder and the classifier head and specifically focus on the features fed to the classifier head and/or modify the classifier weights to mitigate spurious correlations. The two most obvious works are DFR [1] and AFR [2], but there are many others one can find by looking through the papers that have cited these two. This family of works requires its own subsection under Related Works and comparisons in the Experiments section. In its current form, SAS is compared against only two works (LfF and JTT) from 2020 and 2021, respectively. A more comprehensive and detailed literature review and comparison against annotation-free methods should be presented. Furthermore, the paper identifies multiple spurious cues as an important problem but evaluates its method on relatively easy datasets such as Biased-MNIST. I wonder if the proposed annotation method could achieve high accuracy on harder datasets such as UrbanCars. One less significant comment I have is: why do the authors use Unbiased Accuracy (UA) instead of the commonly used Worst Group Accuracy (WGA)?
  • The idea and the proposed direction itself have been studied for a while now (see above). Many studies have linked the problem of mitigating spurious correlations to the simplicity bias phenomenon, where the model learns easy features (spurious features) before learning more structured, complex features (core features). From my perspective, splitting the model into the feature encoder and the classifier head, and then performing several operations to align the outputs of the feature encoders or modify the classifier heads, are all very similar approaches. The novelty claims of this paper would significantly benefit from comparisons against similar methods and, again, more generally, annotation-free methods that have been published since 2021.

[1] Kirichenko, Polina, Pavel Izmailov, and Andrew Gordon Wilson. "Last layer re-training is sufficient for robustness to spurious correlations." arXiv preprint arXiv:2204.02937 (2022).

[2] Qiu, Shikai, et al. "Simple and fast group robustness by automatic feature reweighting." International Conference on Machine Learning. PMLR, 2023.

Questions for authors

See the Critical Weaknesses section. My major concerns are regarding the novelty of the proposed approach, apart from a few small components, such as the use of the cosine classifier. The given problem is well studied in the literature, and similar methods have been proposed but have been neglected in the literature review and experiments. Can the authors present a more comprehensive literature review and provide experimental analysis against other annotation-free methods, specifically focusing on methods that split the model into a feature encoder and a classifier head?

4 comments
AuthorsOpenReview ↗

Feature–Head Methods (DFR, AFR) and Novelty

First , Thank you for the positive comments on clarity and soundness, and for the pointers to related work. We can't update the PDF during the discussion, so the new results are below; the related-work section and these comparisons will be added in the revision.

How SAS differs from DFR and AFR. DFR retrains the last layer on a group-balanced held-out set, so it needs group labels. AFR retrains the last layer on held-out data, with weights computed from a first-stage ERM model. Both are post-hoc, two-stage methods: they work on frozen ERM features and need a held-out split. SAS runs in a single stage during training, so the encoder adapts to the constrained head. CKA between ERM and SAS is 0.82 in early-to-mid layers but 0.48 in the deepest layers. SAS also needs no held-out data, no reweighting and no auxiliary model.

We compared them directly (3 seeds). For AFR and DFR, the ERM features are trained on 80% of the training set, and the other 20% is used for retraining and selection. SAS is our rerun in the same pipeline, which is close to the paper's numbers.

CelebA UACelebA WGACOCO UACOCO WGA
SAS (no labels)93.59±0.0270.67±1.8983.50±0.6968.83±3.81
AFR (no group labels)93.19±0.2868.00±3.2781.47±1.2868.52±2.62
DFR (group labels; reference)95.29±0.4783.97±0.6481.19±0.5373.15±2.62

With the same label budget (no group labels), SAS is above AFR on all four metrics, using one training run. DFR uses group labels and, as expected, has much higher CelebA WGA, although its COCO UA is below SAS.

Novelty. We agree that splitting encoder and head is not new in itself. Our contribution is a specific mechanism that can be analyzed: leverage scores of the batch's dominant right-singular subspace decide which head coordinates to attenuate during training. The revision adds three results that make this concrete, with PP the top-kk projector, MM the selection mask and D=I−(1−α)MD=I-(1-\alpha)M:

  • Prop. A is an exact one-step formula for how the head's alignment with the dominant subspace, ∥Pw∥\|Pw\|, changes.
  • Prop. B is a multi-step bound: alignment is driven down to the part carried by unselected coordinates, so shared task signal stays.
  • Prop. C shows that among coordinate sets of the same size, the top-leverage set gives the largest expected reduction.

The experiments support each part. Selecting the same number of random coordinates does not reduce alignment: head energy in the top-16 subspace on COCO is 0.90 for random selection and 0.33 for SAS. Attenuating the whole subspace instead is worse than SAS in all five settings (on Biased-MNIST, 87.98 / 64.57 / 30.85 vs. 88.85 / 69.80 / 34.32), and hard removal sends CelebA and COCO to chance. DFR, AFR and other last-layer methods have no counterpart to this mechanism or these guarantees.

AuthorsOpenReview ↗

simplicity bias, recent baselines, UrbanCars, WGA

Simplicity bias. We agree this is the right way to look at it, and the revision will discuss it explicitly. Easy spurious features are learned first and dominate the top of the spectrum, and SAS caps how much the head relies on those directions so that weight moves to the rest. We measured head energy in the top-kk subspace against a cosine-only model with the same head (3 seeds). It goes from 0.42 to 0.20 on Biased-MNIST, from 0.98 to 0.84 on CelebA and from 0.44 to 0.33 on COCO.

More recent annotation-free baselines. In addition to JTT and LfF, we ran Spectral Decoupling (the closest single-stage remedy for simplicity bias), BSP and FLAC-B, with the same backbone, epoch budget and checkpoint rule as SAS (3 seeds). Biased-MNIST reports unbiased accuracy and CelebA/COCO report UA. Each baseline is shown with the better of the two λ\lambda/α\alpha values we tried.

MNIST q=.90q=.95q=.99CelebACOCO
SAS88.8569.8034.3293.5983.50
Spectral Decoupling86.8753.7825.1192.9883.77
FLAC-B80.4357.1636.0494.3681.65
BSP90.4476.9742.0393.2683.32

SAS is clearly ahead of Spectral Decoupling on Biased-MNIST (+2.0 / +16.0 / +9.2 points) and on CelebA, and on COCO the two are within one standard deviation. Against FLAC-B, SAS is ahead on Biased-MNIST (q ≤ 0.95) and on COCO. FLAC-B does better on CelebA and BSP on Biased-MNIST, and we will list both as limitations. BSP regularizes the encoder spectrum while SAS constrains the head, so the two could be combined. The related-work section will also cover uLA, SELF, DeFund and SPARE.

UrbanCars. Biased-MNIST is our controlled benchmark with two bias attributes. The main evidence is CelebA and MS-COCO (ResNet-18, two candidate biases each), plus ViT-S/B on COCO (App. K.10). We couldn't finish UrbanCars within the discussion period and will add it to the revision.

UA vs. WGA. We used UA (accuracy averaged uniformly over label × bias cells) because it is the standard metric of the multi-bias protocol we follow (BC-BB, FLAC, GMBM), which let us reuse its setup and baselines. We agree that WGA is stricter and more widely used, so we now report it over all intersectional groups with at least 20 test images. SAS gets 70.67±1.89 on CelebA and 68.83±3.81 on COCO; AFR and DFR are in Part 1, and the group counts are in our reply to Reviewer LdBy. The smallest COCO groups included have only 32–36 test images, so COCO WGA moves by about 3 points per image. The revision will report WGA for every method.

In short, we added a direct comparison with feature–head methods (DFR, AFR) and with three more annotation-free and spectral baselines (Spectral Decoupling, BSP, FLAC-B). We also stated the novelty more clearly: a leverage-selected head update with provable alignment reduction, applied in a single training run. We hope this addresses your concerns about positioning and novelty, and we would be grateful if you would consider raising your score.

Thanks for the response, I will go over it in detail. Quick note on this:

First , Thank you for the positive comments on clarity and soundness, and for the pointers to related work. We can't update the PDF during the discussion, so the new results are below; the related-work section and these comparisons will be added in the revision.

ICLR allows the uploaded PDF to be updated during the rebuttal period, so I do not understand why you state "We can't update the PDF during the discussion".

Thank you for pointing this out, and sorry for the confusion. When we wrote the response, we did not know the PDF could be updated during the discussion period. We have now uploaded a revised PDF, with all new or changed text in blue. Everything promised in our response is now in the paper:

  • Related work on feature–head methods (DFR, AFR, SELF), spectral and simplicity-bias regularizers (Spectral Decoupling, BSP, DeFund) and recent annotation-free methods (uLA, SPARE): Sec. 2, with an extended version in App. N.
  • Comparison with AFR and DFR, including worst-group accuracy: App. M.4, Table 26. The intersectional group counts are in Table 27.
  • Spectral Decoupling, BSP and FLAC-B under the same protocol: App. M.3, Table 25.
  • Simplicity-bias view and head-energy measurements: App. N and App. M.5, Table 28.
  • Novelty / analysis: the alignment results are summarized in Sec. 3.5 and proved in App. E.5 (Props. E.5–E.7). Coordinate selection is compared with whole-subspace attenuation in App. M.2, Table 24.
  • UrbanCars is listed as limitation (4) in App. A.

We would be glad to hear any further comments.

Reviewer_6YbJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an annotation-free method for mitigating spurious correlations by suppressing classifier feature coordinates associated with dominant spectral directions. The authors analyze target and bias information in spectral subspaces and evaluate SAS on several biased benchmarks. Experimental results show improvements over ERM and existing annotation-free debiasing methods. However, whether dominant spectral directions reliably identify bias-related features remains an important question.

Critical strengths

1.Annotation-free mitigation of spurious correlations is an interesting and practically relevant research problem.

2.Analyzing target and bias information in dominant spectral subspaces provides potentially useful insights into the structure of biased representations.

Critical weaknesses

1.Insufficient justification for the proposed suppression strategy. The method directly suppresses classifier feature coordinates associated with dominant spectral directions. However, these directions may encode both spurious and task-relevant information. Although soft suppression may mitigate the loss of useful information to some extent, it does not fundamentally address how to reliably identify bias-related features. The rationale and limitations of this strategy require further justification.

2.Insufficient discussion of multi-attribute bias. The abstract and introduction emphasize scenarios involving multiple spurious attributes, but the method section does not sufficiently explain how SAS handles such scenarios or whether its effectiveness depends on the number and structure of bias attributes.

3.Poor writing and organization. The paper requires substantial revision to improve clarity, logical flow, and technical presentation.

Questions for authors

1.Line 50: What do (N) and (d) represent? Please define these variables when they are first introduced.

2.Lines 152 and 297; Equation (3): The definition of (\hat{z}) at Line 152 and the notation (e_j) at Line 297 are missing or unclear. Equation (3) also appears to contain an indexing error. Please clarify the notation and correct the equation.

3.Punctuation: Colons are used at the ends of some paragraphs where the subsequent text does not appear to introduce a list or explanation. Please check the punctuation throughout the manuscript.

4.Missing tables: Tables 11, 12, and 13 appear to be missing from the manuscript. Please check the completeness of the submitted version.

5.Insufficient methodological explanation: Section 3.2.1 moves directly from the observation of spectral concentration to evaluation metrics without adequately motivating the per-direction analysis or specifying the hypotheses it is intended to test. Section 3.2.2 provides little discussion beyond the figure-caption-level description. Section 3.3 presents an algorithm without sufficient explanation of its rationale and operation. These sections need clearer motivation, interpretation, and methodological details.

3 comments
AuthorsOpenReview ↗

soundness (W1, W2)

Thank you for the review. The soundness concern is fair, and we have tried to answer it with both theory and new experiments. We have uploaded a revised PDF (changes in blue), and we point to the relevant sections below.

W1: is it justified to suppress dominant directions when they also carry task information?

We agree that geometry alone can't separate bias from task information when the two share a subspace. The revision states this in Sec. 3.5 and as limitation (1) in App. A. We don't claim that SAS identifies bias. What we can now show is what SAS does to the head, and why its soft, per-coordinate form keeps task information. Let PP project onto the batch's top-kk right-singular subspace, MM be the selection mask and D=I−(1−α)MD=I-(1-\alpha)M. The three results below are summarized in Sec. 3.5 and proved in App. E.5.

  • Prop. E.5 gives the exact one-step change: ∥PDw∥2−∥Pw∥2=−(1−α2)∥PMw∥2−2(1−α)⟨PMw,P(I−M)w⟩\|PDw\|^2-\|Pw\|^2=-(1-\alpha^2)\|PMw\|^2-2(1-\alpha)\langle PMw,P(I-M)w\rangle. The head's alignment with the dominant subspace goes down at each update, except when the selected and unselected coordinates nearly cancel inside it (as in Reviewer LdBy's example, where the alignment is already zero).
  • Prop. E.6 bounds where repeated updates lead: ∥PDsw∥2≤αs∥Pw∥2+(1−αs)∥P(I−M)w∥2\|PD^sw\|^2\le\alpha^s\|Pw\|^2+(1-\alpha^s)\|P(I-M)w\|^2. Reliance on the dominant subspace is pushed toward the part carried by the unselected coordinates, which SAS leaves alone. That residual is what keeps the target signal that shares the subspace.
  • Prop. E.7: among coordinate sets of the same size, the top-leverage set gives the largest expected reduction of this alignment.

The intuition comes from simplicity bias. An ERM head leans on the few highest-variance directions, which is where easy spurious cues tend to sit. SAS caps that reliance, so the gradient has to use the other directions. We checked this by measuring head energy in the top-kk feature subspace, against a cosine-only model (same head, no suppression; 3 seeds; App. M.5, Table 28):

Biased-MNIST A4A_4CelebA A1A_1COCO A16A_{16}
cosine-only0.42±0.010.98±0.010.44±0.04
SAS0.20±0.060.84±0.040.33±0.02

We also compared SAS with attenuating the whole top subspace (same kk, α\alpha and head; 3 seeds; App. M.2, Table 24). Biased-MNIST reports unbiased accuracy; CelebA and COCO report UA averaged over the two biases. SAS is our rerun in the same pipeline, close to the paper's numbers.

MNIST q=.90q=.95q=.99CelebACOCO
SAS88.8569.8034.3293.5983.50
soft subspace attenuation87.9864.5730.8593.1180.56
hard subspace removal–––50.0050.00

Hard removal sends CelebA and COCO to chance, because it deletes the task signal in that subspace. On COCO, removing only the top two directions already drops target balanced accuracy from 75.84% to 49.92%. Soft attenuation is below SAS in all five settings. This is the situation you describe, and in it the per-coordinate version keeps task information that subspace removal loses.

W2: multiple bias attributes. This is now a paragraph in Sec. 3.4.2 ("Multiple bias attributes"). SAS doesn't need to know the number mm of attributes or which ones they are. If attributes b1,…,bmb_1,\dots,b_m are encoded along directions with variance above σk+12\sigma_{k+1}^2, they all lie in the estimated top-kk subspace. The score ℓj=∑r≤kVjr2\ell_j=\sum_{r\le k}V_{jr}^2 adds up a coordinate's coverage over all dominant directions, so a coordinate is attenuated if it loads on any of them. Correlated attributes also share directions: on CelebA, Lipstick and HeavyMakeup are both decodable from direction 1. What is needed is kk at least the number of dominant spurious directions; we use k=7,26,4k=7, 26, 4 with m=2m=2. Attributes with variance below σk2\sigma_k^2, or encoded nonlinearly, are not covered; this is limitation (2) in App. A.

AuthorsOpenReview ↗

presentation and clarity (W3, Q1–Q5)

These points are all valid, and each is fixed in the revised PDF.

Q1. NN is the mini-batch size and dd the penultimate feature dimension (128 for the Biased-MNIST CNN, 512 for ResNet-18). Both are now defined at first use (Sec. 1), and Sec. 3.1 uses nn for the dataset size to avoid the clash.

Q2. zˉ=1N∑ig(xi)\bar z=\frac1N\sum_i g(x_i) is the batch feature mean, now defined in Sec. 3.1, and eje_j is the jj-th standard basis vector of Rd\mathbb R^d, defined right after Eq. (3). Eq. (3) now reads ℓj=∑r=1kVjr2=ej⊤Pej\ell_j=\sum_{r=1}^{k}V_{jr}^2=e_j^\top Pe_j with P=V:,1:kV:,1:k⊤P=V_{:,1:k}V_{:,1:k}^\top. Sec. 3.4.2 also gives the number of coordinates selected per step: 13/128 on Biased-MNIST, 103/512 on CelebA and 52/512 on COCO.

Q3. We removed the paragraph-final colons that don't introduce a list.

Q4. Tables 11–13 are in the appendix, which was missing from the submitted version (Reviewer LdBy noticed this too), and we apologize for that. The revised PDF includes the full appendix, and the main text now cites each table with its location (App. I). Briefly:

  • Table 11 is the class-balanced spectral audit on MS-COCO.
  • Table 12 has the per-direction probe, MI and η2\eta^2 statistics; this is where both CelebA biases show up on direction 1.
  • Table 13 is layer-wise CKA between ERM and SAS: 0.82 in early-to-mid layers and 0.48 in the deepest ones, so SAS mostly changes the high-level features.

Q5. Sec. 3.2.1 now begins by stating what it tests: (H1) under ERM, spurious information concentrates in the top singular subspace, and (H2) how strongly bias is enriched there relative to the label depends on the dataset. The per-direction analysis then checks this one direction at a time. Sec. 3.2.2 now explains that the ablation asks whether this information is localized there. Removing the top-16 directions lowers bias-probe accuracy by 32.4/44.0 points, while removing 16 random directions has no effect.

Sec. 3.3 is now a step-by-step walk-through of Algorithm 1: center the batch features, take the top-kk right singular vectors, compute the leverage scores, then scale the selected head columns by α\alpha after the optimizer step (so the gradient update itself is unchanged). It links these steps to Props. E.5–E.7.

To sum up, for W1 we now have a statement of what SAS does to the head's subspace alignment, plus experiments where the per-coordinate design keeps task information and subspace removal does not. For W2 there is an explicit coverage condition, and each clarity issue is fixed in the revised PDF. We hope this changes your view of the paper, and we would be grateful if you would reconsider your score. We are glad to answer any further questions.

Thans for the authors' response and the additional experimental analyses. However, I still have concerns about whether the experimental evidence sufficiently supports the claims made in the paper. For example, the authors introduce an analysis of the classifier's head energy in the top-feature subspace. However, this analysis primarily measures how the classifier's reliance on the dominant subspace changes after applying SAS. While it provides evidence that SAS reduces such reliance, it does not directly establish that the model's reliance on spurious information is reduced. In other words, the current experiments mainly demonstrate changes in subspace reliance and target performance, but do not clearly verify the claimed reduction in bias reliance.

Moreover, several of my original concerns remain unresolved. e.g. 1) since dominant spectral directions do not necessarily correspond to bias-related features and may also contain task-relevant information, why should suppressing these directions effectively mitigate bias without compromising task performance? 2) when multiple bias attributes are present, why can we expect all high-variance bias-related information to be captured by the top-k subspace? These questions are central to understanding the rationale and applicability of the proposed suppression strategy, but the current response does not provide sufficiently convincing answers.

Also, the readability of the manuscript remains a concern. The presentation is still difficult to follow, which makes it challenging to assess the experimental design and determine whether the reported results adequately support the proposed method and its claims. I encourage the authors to further clarify the underlying rationale, better organize the presentation, and improve the overall quality of the manuscript.

Since these fundamental concerns remain unresolved, I will maintain my current score.

Reviewer_LdByOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Spectral Alignment Suppression (SAS), a single-model method for mitigating spurious correlations without per-sample bias labels in training. It estimates the leading right-singular subspace of centered minibatch features, computes coordinate leverage scores, and attenuates high-leverage columns of a cosine classifier after gradient updates. The spectral audit identifies bias-dominated directions on Biased-MNIST and shared target–bias concentration on CelebA and COCO. Reported experiments show substantial gains over ERM and improvements over JTT/LfF in most evaluated settings. The supplied PDF references appendices containing the formal theory and additional experiments, but those appendices were not available for this assessment.

Critical strengths
  1. The method is simple and avoids an auxiliary biased model. Its use of the model's own feature geometry is practically appealing. The paper carefully acknowledges that the dominant subspace is a dataset-dependent risk prior rather than a bias-only identifier, particularly on CelebA and COCO (pp.2–4).

  2. The main tables show meaningful improvements over ERM across the biased benchmarks, with three-seed statistics. For example, COCO Sports improves from 70.86 to 83.77 UA and from 64.61 to 83.51 BCA. Table 2 includes cosine-only and random-suppression controls on Biased-MNIST, which help distinguish spectral selection from a generic classifier change. The explicit statement of where bias information is used in validation and evaluation is also useful.

Critical weaknesses
  1. The justification for the specific intervention remains incomplete in the available material. High subspace leverage identifies coordinates, not a bias-only subspace, and diagonal attenuation need not decrease the classifier's actual projection onto that subspace. For example, with top vector v=(0.8,0.6) and classifier row w=(0.75,-1), w dot v is initially zero but becomes -0.3 after halving the first, higher-leverage coordinate. This does not contradict contraction of a leverage-weighted coordinate energy; it limits what such contraction establishes. The main paper does acknowledge shared target–bias information. Matched cosine-only/random controls across CelebA and COCO, a direct subspace-attenuation comparison, and clearer diagnostic evidence are needed to establish why coordinate selection preserves target information.

  2. The comparison framing needs correction. The original FLAC paper explicitly describes bias-label-unaware mitigation (Sarridis et al., https://arxiv.org/abs/2304.14252), whereas pp.1,3 and Tables 2–4 classify it as requiring per-attribute bias labels. Unless the evaluated implementation changes this requirement, FLAC belongs in the annotation-free comparison; its CelebA Lipstick UA (95.40 versus SAS 94.95) and HeavyMakeup BCA (87.20 versus 85.82) affect the best-in-category claims. In addition, per-attribute UA/BCA do not establish performance on the worst intersectional group. The main-paper results do not resolve this distinction, and the referenced additional evaluation was unavailable.

Questions for authors
  1. Does your FLAC implementation use bias labels? If not, please correct its categorization and the corresponding best-annotation-free claims. Please also position SAS against low-rank debiasing, Batch Spectral Penalization and Spectral Decoupling, distinguishing motivation from algorithmic equivalence.

  2. What connects the leverage-weighted quantity in the analysis to actual classifier subspace alignment and selective bias reduction? Can you compare coordinate attenuation with direct top-subspace attenuation and report matched cosine-only/random controls on all main datasets? How are percentile ties handled, and how many coordinates are actually selected?

  3. Please report label-by-bias intersectional group counts and worst-group performance, and clarify validation/checkpoint selection on CelebA. How should the increases in both CelebA bias MI estimates in Table 1 be reconciled with the reduced bias-probe accuracy? Please include uncertainty rather than interpreting either estimate as definitive information removal.

3 comments
AuthorsOpenReview ↗

theory, subspace attenuation, selection details

Thank you for the detailed review. Your counterexample is right, and working through it gave us a cleaner analysis than the one in the paper. We have uploaded a revised PDF (changes in blue) and point to the relevant sections below. Also, apologies that the appendix was missing from the version you received; the revised PDF includes it.

Q1: from leverage to alignment. We agree that Prop. 2 (contraction of ∑jℓj∥W:,j∥2\sum_j \ell_j\|W_{:,j}\|^2) does not imply contraction of ∥Pw∥\|Pw\|. Since ∥Pw∥2=∑jℓjwj2+∑i≠jPijwiwj\|Pw\|^2=\sum_j \ell_j w_j^2+\sum_{i\neq j}P_{ij}w_iw_j, SAS only controls the diagonal part, and your example works through the cross-terms. Write MM for the selection mask, D=I−(1−α)MD=I-(1-\alpha)M, a=PMwa=PMw and b=P(I−M)wb=P(I-M)w. The three results below are summarized in Sec. 3.5 ("Effect on subspace alignment") and stated and proved in App. E.5.

Prop. E.5 (one step). ∥PDw∥2−∥Pw∥2=−(1−α2)∥a∥2−2(1−α)⟨a,b⟩\|PDw\|^2-\|Pw\|^2=-(1-\alpha^2)\|a\|^2-2(1-\alpha)\langle a,b\rangle. Alignment goes down unless the selected and unselected parts nearly cancel inside the subspace, i.e. ⟨a,b⟩≤−1+α2∥a∥2\langle a,b\rangle\le-\tfrac{1+\alpha}{2}\|a\|^2. Your example sits on this boundary: Pw=0Pw=0 (a=−ba=-b), so there is no alignment left to remove, and the identity gives 0.09=0.320.09=0.3^2.

Prop. E.6 (several steps). By convexity in α\alpha, ∥PDsw∥2≤αs∥Pw∥2+(1−αs)∥P(I−M)w∥2\|PD^sw\|^2\le\alpha^s\|Pw\|^2+(1-\alpha^s)\|P(I-M)w\|^2. Alignment is pulled toward what the unselected coordinates carry, and it strictly decreases whenever ∥Pw∥>∥P(I−M)w∥\|Pw\|>\|P(I-M)w\|.

Prop. E.7 (why leverage). For an isotropic row, E∥PDw∥2=ν2(k−(1−α2)∑j∈Hℓj)\mathbb E\|PDw\|^2=\nu^2\big(k-(1-\alpha^2)\sum_{j\in H}\ell_j\big), and the fixed point of Prop. E.6 has residual coverage k−∑j∈Hℓjk-\sum_{j\in H}\ell_j. Over all sets with ∣H∣=m|H|=m, both are optimized by the top-mm leverage scores.

Each proof takes a few lines, and we checked all three numerically on 20,000 random instances. They concern the unnormalized WW that SAS actually modifies; we don't claim monotonicity for the cosine-normalized prototype (this scope is stated in App. E.5).

To see whether this holds in trained models, we measured A16=∥WP16∥F2/∥W∥F2A_{16}=\|WP_{16}\|_F^2/\|W\|_F^2, the head energy in the top-16 test-feature subspace (App. M.5). We compared SAS with a cosine-only model (same head, no suppression) and with random suppression of the same number of coordinates. On COCO it is 0.44 for cosine-only, 0.90 for random and 0.33 for SAS; on CelebA it is 1.00 / 0.98 / 0.91. Random selection does not lower it, which fits Prop. E.7.

Direct subspace attenuation. We ran it with the same kk, α\alpha and head (3 seeds; App. M.2, Table 24). Biased-MNIST reports unbiased accuracy; CelebA and COCO report UA averaged over the two biases. SAS here is our rerun in the same pipeline as the baselines, and it is close to the paper (88.85 / 69.80 / 34.32 vs. 88.22 / 67.63 / 34.89 on Biased-MNIST).

MNIST q=.90q=.95q=.99CelebACOCO
SAS (coordinates)88.85±1.0369.80±1.6534.32±1.1093.59±0.0283.50±0.69
soft subspace W(I−(1−α)P)W(I-(1-\alpha)P)87.98±1.0764.57±0.5030.85±0.3393.11±0.3280.56±1.53
hard subspace W(I−P)W(I-P)–––50.0050.00

Scaling the whole subspace hits the in-subspace component of every coordinate, including the target signal that lives there too. Hard removal drops to chance, and the soft version is unstable on COCO (last epoch 59.67±13.87 vs. 81.23±1.30 for SAS). SAS scales only the highest-leverage coordinates, so at its fixed point P(I−M)w≠0P(I-M)w\neq0, and that residual is what keeps the shared target signal.

Ties and number of selected coordinates (now in Sec. 3.4.2). Scores are continuous, so exact ties have probability zero. Every coordinate at or above the 100(1−τ)100(1-\tau)-th percentile is selected, so a tie at the threshold would be included. Per step we select 13/128 coordinates on Biased-MNIST (k=7k=7), 103/512 on CelebA (k=26k=26) and 52/512 on COCO (k=4k=4). The selection is also fairly persistent, which is the setting of Prop. E.6 (App. M.5). Mean Jaccard overlap between consecutive steps is 0.65 on CelebA, 0.33 on COCO and 0.28 on Biased-MNIST, against 0.11, 0.05 and 0.05 for random sets of the same size.

FLAC, the other baselines, WGA, checkpoint selection and MI are in Part 2.

AuthorsOpenReview ↗

baselines, groups, protocol, MI

FLAC. You're right, our description was off. FLAC (Sarridis et al., Sec. 3.2/4.3) has two variants. The one in our tables trains its bias-capturing network with protected-attribute labels, while FLAC-B uses a vanilla model and no attribute labels. The revision describes this correctly in Sec. 1, Sec. 2 and the captions of Tables 2–4. Because the variant we evaluate uses attribute labels, it stays in the supervised group, and Sec. 5.3 states where it is ahead of us (Lipstick UA 95.40 vs. 94.95, HeavyMakeup BCA 87.20 vs. 85.82). We also ran FLAC-B as an annotation-free baseline (table below).

Low-rank debiasing, BSP, Spectral Decoupling. These overlap with SAS in motivation but not as algorithms (Sec. 2 and App. N). DeFund (CVPR'24) also links bias to low effective rank, but it rank-regularizes a self-supervised biased encoder and then upweights bias-conflicting samples, so it is a two-stage pipeline. BSP penalizes the top singular values of the features, which changes the encoder's objective. Spectral Decoupling penalizes ∥logits∥2\|\text{logits}\|^2, which is isotropic in feature space. SAS uses a single model and limits only the head's use of the dominant subspace.

We ran these in the same pipeline as SAS (3 seeds, two hyperparameter values each; App. M.3, Table 25):

MNIST q=.90q=.95q=.99CelebA UACOCO UA
SAS88.8569.8034.3293.5983.50
FLAC-B (α=1e3 / 1e4)80.4357.1636.0494.36 / 93.5481.65 / 80.99
SD (λ=0.1 / 0.01)86.8753.7825.1192.94 / 92.9883.45 / 83.77
BSP (λ=1e-4 / 1e-3)90.4476.9742.0393.26 / 92.2683.32 / 83.07

SAS is ahead of SD on all three Biased-MNIST settings and on CelebA, and ahead of FLAC-B on Biased-MNIST (q ≤ 0.95) and COCO. On COCO, SD and SAS are within one standard deviation. FLAC-B does better on CelebA and BSP on Biased-MNIST; both are now listed as limitation (6) in App. A. BSP works on the encoder spectrum and SAS on the head, so combining the two seems worth trying.

Intersectional groups. Train/test counts per group (label, b1b_1, b2b_2), also in Table 27 (App. M.4):

group000001010011100101110111
CelebA (Male, Lipstick, Makeup)14792/64061216/55913085/572653541/2284058746/2507070/25271/11398/41
COCO (gender, sports, kitchen)2507/108924/34958/3620/13212/125787/321512/6223/1

We compute WGA over groups with at least 20 test images. Two COCO groups have one test image each, and the smallest remaining ones have 32–36, so a single image moves COCO WGA by about 3 points. SAS gets 70.67±1.89 on CelebA and 68.83±3.81 on COCO. Table 26 (App. M.4) gives UA and WGA for SAS, AFR, DFR and all the baselines above; the AFR and DFR numbers are also in our reply to Reviewer 2fHp. On WGA, SAS is above AFR on both datasets, while on COCO several baselines are above SAS by about two test images, and on CelebA FLAC-B and DFR are higher.

CelebA checkpoint selection (now in App. C). We use the GMBM protocol (70/30 split) for all methods. The checkpoint is picked by average target accuracy on the held-out split, without bias or group labels. Backbone, epochs, split and selection rule are the same for SAS and every baseline.

MI vs. probe accuracy (Table 1). Agreed, the two disagree on CelebA. Both are unconditional, and Lipstick and HeavyMakeup are strongly correlated with gender, so more label information also raises the unconditional bias MI. The ratio MI(y)/MI(b) barely moves (1.36 → 1.36 and 1.70 → 1.73). We added a conditional check that decodes the bias within each gender class from the top-16 subspace (3 seeds; App. M.5). HeavyMakeup decodability drops from 75.23±2.72 (cosine-only) to 70.91±3.09 (SAS), Lipstick is unchanged (78.45 → 78.27), and conditional KSG MI is also unchanged (0.05 / 0.08). We rephrased the claim in Sec. 3.2.3. On CelebA, SAS changes how the head uses the representation (lower A16A_{16}, Part 1); it does not remove bias information, and neither estimate should be read as showing that.

Thanks again for the careful questions. We hope the above answers them, and if so, we would appreciate it if you would reconsider your score. We are happy to follow up on anything during the discussion.

Thank you for restoring the appendices in the revised PDF and providing thorough theoretical and empirical follow-ups.

Several key concerns from my initial review are well addressed in the revision:

  1. Subspace Alignment Analysis & Direct Subspace Comparison: Propositions E.5–E.7 cleanly resolve the gap highlighted by the counterexample by decomposing Pw=a+bPw = a + b and showing that repeated diagonal coordinate attenuation drives alignment toward the unselected residual P(I−M)wP(I-M)w rather than zero. Together with Table 24 (where hard subspace removal collapses CelebA/COCO to 50% and soft subspace attenuation underperforms coordinate selection), this provides a coherent explanation for why selective coordinate decay preserves shared target signal. I also appreciate the explicit scope statement in Appendix E.5 noting that monotonicity applies to the unnormalized rows of WW rather than the cosine-normalized prototypes w^c\hat{w}_c.
  2. FLAC Framing & CelebA MI Audit: Clarifying that the main tables evaluate the attribute-labeled variant of FLAC while adding label-free FLAC-B in Appendix M.3 resolves the categorization issue. Likewise, the gender-conditional probe and KSG MI check in Appendix M.5 sensibly explain the unconditional MI increase in Table 1, and the revised claim in Section 3.2.3—that SAS reduces head reliance (AkA_k in Table 28) on CelebA rather than removing bias information from the representation—is much better supported.

Before finalizing my score, could you clarify two remaining points from the revised results?

  1. Missing Matched Accuracy Controls on CelebA and MS-COCO: In Q2, I asked for matched ERM + Cosine (cosine head without suppression) and Random Suppression (cosine head with random coordinate decay at the same τsas\tau_{\text{sas}} and α\alpha) controls across all main datasets. While Appendix M.5 reports their subspace energy AkA_k and Table 14 ablates SAS (SVD only) on MS-COCO, the actual classification metrics (UA / BCA / WGA) for ERM + Cosine and Random Suppression are still missing on CelebA and MS-COCO. Since you already trained those models to compute AkA_k in Section M.5, could you share their UA, BCA, and WGA on CelebA and MS-COCO so we can isolate how much of the gain over vanilla ERM comes from the cosine classifier versus SVD-guided selection on the real-image benchmarks?
  2. Positioning Against Stronger Single-Stage Baselines & WGA: The new comparisons in Tables 25 and 26 are very informative, though they temper the empirical picture: single-stage annotation-free baselines outperform SAS in several regimes—BSP leads substantially across all Biased-MNIST severities (e.g., 76.97 vs. 69.80 at q=0.95q=0.95; 42.03 vs. 34.32 at q=0.99q=0.99), FLAC-B leads on CelebA in both UA (94.36 vs. 93.59) and WGA (73.33 vs. 70.67), and Spectral Decoupling (λ=0.01\lambda=0.01) edges out SAS on MS-COCO UA (83.77 vs. 83.50) and WGA (74.56 vs. 68.83). Given these results, where do you view SAS as offering the clearest practical advantage over simple logit/spectrum penalties like SD or BSP?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.