acceptodds
Withdrawn from ICLR 2027

UNEVEN NORMAL COVERAGE PERTURBS RESIDUAL ANOMALY METRICS: A SPECTRAL MECHANISM STUDY

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Residual anomaly detectors rely on reference statistics from normal data, making global covariance-based scores vulnerable to unequal coverage of valid normal modes. We present a mechanistic study linking uneven normal-reference coverage to residual-metric changes and target-specific false positive inflation. By isolating the metric pathway, we use a two-mode mixture model to separate within-mode covariance shifts from between-mode mean differences, along with an exact common-basis decomposition to attribute score changes to specific spectral directions. In the primary DINOv3 ViT-B/16 representation, positive net attribution is heavily concentrated in leading spectral directions. Controlled interventions—including sample-matched deletions, reweighting, and spectral adjustments—confirm that these shifts stem from composition and spectral location rather than sample count. As a mechanism-inspired probe, Head-Calibrated Mahalanobis (HCM) mitigates conditional false-positive inflation under reduced coverage, demonstrating that the identified failure mode is actionable. Our work provides a precise mechanistic account of coverage sensitivity in residual anomaly detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Anomaly Detection, Mahalanobis Distance, Uneven Normal Coverage, Majority-Reference Bias, Spectral Calibration

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_Sbc4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the problem where one normal mode may have only a few reference samples. Normal samples from this data-rare mode then receive higher anomaly scores under the global covariance metric. More of them exceed the threshold calibrated at full coverage and lead to increased false positives. The authors propose a two-mode mixture model to separate within-mode covariance shifts from between-mode mean differences. They decompose this score increase along the eigenvectors of the full-coverage covariance. Most of the increase comes from the leading eigenvectors in DINOv3 ViT-B/16. On MVTec AD they compare targeted deletion with matched random deletion and reweighting to distinguish composition from sample count. They also design Head-Calibrated Mahalanobis (HCM) based on these findings to reduce this false-positive inflation at the cost of lower detection performance.

Critical strengths

The problem studied in this paper is valuable for anomaly detection. Normal reference sets can contain several valid modes such as different poses or illumination conditions. These modes can have very different numbers of samples. The resulting false positives on data-rare modes can be hidden by aggregate detection metrics such as image AUROC. The authors split the score increase behind these false positives into exact contributions from each covariance eigenvector. These contributions sum to the total increase and show which eigenvectors drive it.

Critical weaknesses

W1. Inconsistent reporting of the main results. At q = 0.5 the HCM I-AUROC is 0.9061 in Table 10 but 0.925 in Table 11. Table 11 only recalibrates the threshold while I-AUROC depends solely on the ranking of image scores. Mahalanobis has the same I-AUROC in both tables (0.9514 vs 0.951) so the two HCM rows differ in their scores and not only in the threshold. This undermines the claim in Section 6 that HCM helps mainly at fixed thresholds. I am also confused by Appendix A because a holdout reanalysis gives different estimates from the selected primary results. It is unclear which set supports the claims and the primary gains lack uncertainty estimates.

W2. Dependence of the leading-rank taper on the residual scale. Eqs. (9)–(10) imply that multiplying all residuals by ss scales the contribution of eigenvector ii by s2−4γis^{2-4\gamma_i}. With the taper this factor varies across eigenvectors and can change the ranking of image scores. Without the taper (β=0\beta = 0) it is a common factor. The ratio wiHCM/wibase=λ~i2γβtiw_i^{\mathrm{HCM}}/w_i^{\mathrm{base}} = \tilde\lambda_i^{2\gamma\beta t_i} stated after Eq. (10) exceeds one only when λ~i>1\tilde\lambda_i > 1. The residual scale thus decides whether the taper raises or lowers the leading-rank weights. I could not find how ep(x)e_p(x) in Section 3 is constructed or normalized. This weakens the validity of the HCM results.

Questions for authors

Q1. Do Tables 10 and 11 use the same HCM configuration and the same image scores? I suggest that the authors report a matched comparison in which the fixed and recalibrated thresholds are applied to the same scores.

Q2. How are the residuals ep(x)e_p(x) in Section 3 constructed and normalized? I suggest that the authors rescale the same residuals by several factors ss and check whether the reduction in target ΔFPR\Delta\mathrm{FPR} persists. For each ss the threshold can be calibrated at q=1q = 1 and reused at reduced coverage.

Q3. Table 4 reports only the differences between targeted deletion and each control. If its targeted values match Table 3 then matched random deletion gives a target ΔFPR\Delta\mathrm{FPR} of 0.179/0.374/0.496 at the three reduced coverages. This is 67%–79% of the targeted effect while fixed-support reweighting gives only 0.016/0.058/0.155. I suggest that the authors report the absolute target ΔFPR\Delta\mathrm{FPR} of each control.

Q4. Does the leading 15% band in Table 1 cover the same ranks as the first-15% HCM head in Table 2? Both tables report fixed-residual attribution on DINOv3 ViT-B/16 as category–target–seed means. Yet the head contribution is 123.28 to 245.30 in Table 1 and 5550.17 in Table 2.

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

This paper studies how uneven coverage of normal samples affects residual anomaly scores and conditional false positives. The authors use a two-mode covariance model and a fixed-residual spectral decomposition to analyze score changes. They also introduce Head-Calibrated Mahalanobis (HCM), which modifies the global spectral weights and applies an additional exponent taper to leading directions. Experiments show that HCM can not only reduce target FPR inflation under fixed thresholds, but also reduces anomaly detection performance.

Critical strengths
  • The paper studies a clear problem in residual anomaly detection.

  • The fixed-residual analysis is well defined.

  • The experiments include different coverage levels, matched random deletion, fixed-support reweighting, and spectral intervention controls. These experiments can provide useful evidence about the sensitivity of residual metrics to the data.

Critical weaknesses
  • It seems that the motivation is not well-supported. To my understanding the results shown in the paper can suggest a substantial effect from reducing sample support, but it remains unclear how much of the observed failure comes from changes in mode proportions rather than finite-sample covariance estimation. The current experiments do not fully separate these effects.

  • The evidence for a specific leading-direction mechanism may be insufficient. For example, Table 2 shows that residual energy is highly concentrated in the leading spectral directions. The leading third contains an energy of 9240.80, while the middle and tail thirds contain only 12.41 and 7.78. Why this happens? Alternatively, an energy-normalized attribution analysis or other controls might be better to support the proposed spectral explanation.

  • The contribution of the additional head taper seems limited as shown in Table 10, which could raise questions about the necessity of the head taper and the choice of its parameters.

  • Besides, the evaluation can be limited in generality and statistical support. For example, there lacks uncertainty intervals and the consistent threshold calibration protocol. More consistent comparisons are needed.

Questions for authors

Please see the questions and concerns in the critical weaknesses section.

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

This paper investigates how uneven normal-reference coverage affects residual-based anomaly detection through changes in covariance-based scoring. It introduces a spectral attribution analysis to explain the relationship between reference composition, anomaly scores, and conditional false-positive rates. The authors further propose Head-Calibrated Mahalanobis (HCM) to mitigate coverage-induced false positives. Experiments across multiple datasets and feature backbones reveal coverage sensitivity, but also substantial trade-offs between robustness and anomaly discrimination.

Critical strengths
  1. Interesting problem perspective. The paper investigates an important but relatively underexplored issue in anomaly detection, namely the sensitivity of normal-reference statistics to uneven coverage of valid normal modes.

  2. Detailed analytical framework. The spectral decomposition and controlled coverage interventions provide useful observations about how reference composition influences covariance-based anomaly scoring.

Critical weaknesses
  1. Limited methodological novelty and significance. The main analysis largely builds upon established covariance decomposition, Mahalanobis distance, and spectral regularization techniques. Although the observations are interesting, the fundamental methodological contribution beyond applying these mathematical tools to anomaly detection remains unclear.

  2. Substantial degradation in anomaly detection performance. HCM reduces conditional false-positive inflation but significantly sacrifices anomaly discrimination. In Table 10, Image AUROC decreases from 0.9514 to 0.9061, while TPR drops from 0.9697 to 0.8045. The authors should justify whether the improvement in conditional FPR is practically meaningful given these substantial losses.

  3. Poor generalization across feature representations. Table 6 shows that HCM reduces Image AUROC from 0.971 to 0.542 on DINOv2 ViT-B/14. On DINOv3 ViT-S/16, HCM even increases conditional FPR inflation under stronger coverage reduction. These results raise serious concerns about the robustness and generalizability of the proposed spectral intervention.

  4. Insufficient independent contribution of the spectral taper. Table 10 indicates that HCM without the proposed taper already achieves substantial robustness improvements while preserving much better detection performance. The authors should better justify the necessity of the head-specific taper and demonstrate a favorable robustness–accuracy trade-off.

  5. Limited statistical evidence and experimental reliability. Appendix A states that the paired observations for the primary results are unavailable, preventing uncertainty intervals from being reconstructed. This considerably weakens the statistical support for the central experimental claims. More complete paired results and uncertainty estimates are needed.

Questions for authors

See the weaknesses part.

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

Related papers

Discussion (0)

Sign in to comment.