Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Assessing the Noise Robustness of Class Activation Maps: A Framework for Reliable Model Interpretability

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A heatmap-based explanation is robust only if it stays stable when noise leaves the predicted class unchanged and visibly reorders when the prediction flips; the paper scores both with a Consistency × Responsiveness metric.

desk verdict A useful robustness metric for CAMs with a broad empirical sweep, but the severity table and the in-sample AUC need fixing before the rankings can be trusted. read the letter →

arxiv 2508.18154 v1 pith:Y6B47KDP submitted 2025-08-25 cs.CV cs.AIcs.LG

Syamantak Sarkar , Revoti P. Bora , Bhupender Kaushal , Sudhish N George , Kiran Raja This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords classactivationmapsexplanationrobustnessRank-BiasedOverlapconsistencyresponsivenessnoiseperturbationmodelinterpretabilityCAMbenchmarking
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 proposes a way to score how well Class Activation Maps — the heatmaps showing which image regions drove a model's decision — survive input noise. The central claim is that a robust CAM must do two things at once: keep its segment rankings stable when a perturbation leaves the predicted class alone (consistency), and reorder its highlighted regions when the perturbation actually changes the prediction (responsiveness). The authors compress both into a single number, Robustness = Consistency × Responsiveness, computed with Rank-Biased Overlap on segment rankings before and after perturbation. Across four architectures, five datasets, and eight perturbation types, the metric ranks GradCAM++ first and EigenCAM and AblationCAM last. If the metric is right, it gives practitioners a model- and noise-agnostic tool for choosing an explanation method that will not silently mislead under real-world distortions.

What carries the argument

Rank-Biased Overlap (RBO) with persistence parameter p = 0.9 — a top-weighted similarity measure for indefinite rankings — applied to segment-wise mean CAM intensities. Images are first cut into QuickShift superpixels; each CAM is averaged per segment to give a ranked list; RBO compares the clean and perturbed rankings. RBO does the work the paper claims normalized ℓ1 distance cannot: it detects semantic reordering of salient regions rather than raw intensity shifts. The Responsiveness term repurposes the same RBO values as features in a binary classifier whose AUC measures how well a CAM's rank change tracks the model's decision change.

What would settle it

Recompute Responsiveness with 5-fold cross-validation or a held-out split of the image-level RBO samples and check whether GradCAM++ still leads and EigenCAM still trails; if cross-validated AUCs collapse toward 0.5 or the ordering changes, the published rankings are fitting artifacts. A second check: shuffle the RBO values against the class-change labels for one CAM method — a valid metric should give near-chance responsiveness AUC (~0.5) and a lower product score, whereas a metric that rewards any stable signal would not.

Watch

Extended reading notes

Core claim

The paper's core assertion is that prior robustness checks conflate two failure modes: a CAM can be unstable even when the model still predicts the same class, or it can be immobile — producing nearly identical heatmaps — even when the model's prediction has genuinely changed. Consistency is defined as the median Rank-Biased Overlap between segment-importance rankings of clean and perturbed images, restricted to perturbations that leave the predicted class unchanged; Responsiveness is the AUC of a linear classifier that must predict, from the RBO value alone, whether the class changed. The Robustness Metric is their product. Over six CAM methods, four models (ResNet50, VGG19, InceptionV3, Vi

Load-bearing premise

The metric's responsiveness half stands on an AUC computed by fitting a classifier to the same RBO values it is then scored on, so any in-sample optimism in that AUC would inflate the robustness rankings built from it.

Editorial extensions

If this is right

  • Practitioners can use the Consistency × Responsiveness product as a selection criterion for CAM methods in noise-prone applications; the paper's sweep suggests GradCAM++ is the most reliable default among the six tested.
  • EigenCAM and AblationCAM should be treated with caution in high-stakes settings, since their apparent stability reflects unresponsiveness to real prediction changes rather than genuine robustness.
  • Transformer-based models (ViT) show markedly higher variance in CAM robustness scores, so reliability expectations set on convolutional models do not transfer to ViT explanations.
  • The framework's rankings are nearly invariant to design choices: segmentation method (QuickShift, SLIC, Felzenszwalb) and rank-correlation metric (RBO, Kendall's τ, Spearman's ρ) leave the ordering of CAM methods essentially unchanged.
  • Stochastic CAMs (SmoothGrad, VarGrad, CAPE) score close to GradCAM++ under the same metric, indicating the framework extends beyond deterministic methods.
  • The responsiveness component is what separates robust from apparently stable methods: EigenCAM scores highest in low-noise regimes where consistency dominates, then collapses once predictions start changing.

Reading between the lines

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

  • The responsiveness AUC is computed in-sample with no held-out split; if it were recomputed with cross-validation, absolute scores would likely drop, and a direct check of whether the CAM ranking survives would settle whether the published order is a fitting artifact.
  • The product form encodes an implicit preference for balance: a perfectly stable but unresponsive method scores zero, and so does a perfectly responsive but unstable one — a design choice that could be debated against additive or weighted combinations.
  • The same consistency/responsiveness decomposition could be lifted to other explanation families such as LRP, SHAP, or attention maps, turning the framework into a general explanation-robustness benchmark rather than a CAM-specific one.
  • Because RBO's persistence parameter p weights top-ranked segments, lowering p would emphasize the single most salient region — a variant that could change rankings in medical-imaging use cases where the leading region matters more than the full ordering.
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 / 5 minor

Summary. The paper proposes a framework for assessing the noise robustness of Class Activation Maps (CAMs). For each image, CAM method, and perturbation type, the pipeline segments the image, computes CAMs for clean and perturbed inputs, aggregates CAM intensities per segment, and forms ranked lists of segments. Rank-Biased Overlap (RBO) measures ranking similarity. The Robustness Metric is defined as Consistency × Responsiveness, where Consistency is the median RBO over perturbations that do not change the predicted class and Responsiveness is the AUC of a classifier that predicts whether the class changed from the RBO score. The authors evaluate six CAM methods across four architectures and five datasets, report GradCAM++ as the most robust and EigenCAM/AblationCAM as the least robust, and include ablations over segmentation methods, rank correlation metrics, perturbation severity, probabilistic CAMs, and adversarially trained models.

Significance. If validated, the proposed framework would be a useful, model- and CAM-agnostic benchmarking tool for explanation robustness. The paper's strengths are its breadth (four architectures, five datasets, eight perturbation families, six CAMs), the clear motivation for RBO over ℓ1-style intensity differences, and the segmentation and rank-metric ablations, which show that qualitative rankings are stable under those design choices. However, the central metric's reliability depends on the responsiveness estimator and on the consistency of the severity-level data. The current version has internal inconsistencies, especially in the severity ablation, and the responsiveness estimate has an unaddressed sample-sparsity problem. These issues are load-bearing because they affect the validity of the reported CAM rankings.

major comments (3)
  1. [Table 13 / Sec. 4.3.3] Table 13 is internally inconsistent and contradicts Tables 4–7. For example, on ImageNet/ResNet-50 the Gaussian low-severity row gives EigenCAM a Robustness score of 0.650 ± 0.03, whereas Table 4 reports EigenCAM Gaussian = 0.215 ± 0.02 for the same condition. Moreover, for every natural perturbation the medium and high rows are numerically identical (e.g., Gaussian Medium = High for all CAM methods), and for Poisson, Table 3 explicitly states 'No parameters (applied directly),' yet Table 13 assigns Low/Medium/High severity levels to Poisson. This makes the severity ablation uninterpretable and undermines the severity-dependent conclusions. The authors need to rerun or re-report these results with actual perturbation parameters and, for each cell, the number of images whose class changed.
  2. [Eq. (4) / Algorithm 1, lines 24–25] Responsiveness is computed as the AUC of a linear classifier trained and evaluated on the same RBO/class-change data, with no cross-validation and no minimum-sample safeguard. For a single feature, in-sample AUC equals the empirical concordance probability, so the no-CV concern is modest. The load-bearing problem is sparsity: at low severity (Table 13) very few images change class, so the AUC is undefined if there are zero positives and extremely noisy if there are only a handful; similarly, Consistency in Eq. (3) is a median over potentially very few unchanged images at high severity. Yet finite Robustness scores are reported for such cells. Please report class-change counts per cell, use a cross-validated or closed-form concordance estimate with confidence intervals, and exclude or flag cells below a minimum positive/negative sample size. Without this, the product in Eq. (2) is not a r
  3. [Sec. 2.5 / Sec. 4.1] The conclusion that 'GradCAM++ is the most robust method' is obtained from the proposed Robustness Metric itself, and no external criterion is supplied to validate that Consistency × Responsiveness equals robustness of CAM explanations. Figures 8–9 are qualitative. Because the product is a definitional choice, the ranking is not circular in the logical sense, but the external validity of the framework remains unestablished. A concrete check would be to compare the metric's ranking against an established interpretability-quality measure (e.g., ROAR/LEAF-style fidelity drop or a human agreement study) on a subset of the data, or to show that the metric predicts performance in a downstream explanation-reliability task. At minimum, the paper should state explicitly that the claims are limited to the proposed operationalization.
minor comments (5)
  1. [Sec. 3 vs. Sec. 4.4] Section 3 states that 1000 images are selected per dataset, but Section 4.4 and Table 14 compute runtime for '100 images.' Please reconcile the number.
  2. [Abstract / Sec. 1.1] The contributions and abstract mention 'seven diverse datasets,' but only five datasets are actually used (OxfordPets, Dogs-vs-Cats, ImageNet, Melanoma, Caltech).
  3. [Algorithm 1, line 12] Line 12 refers to 'ynoisy,' which is not defined; line 9 defines 'yper.' Please use consistent variable names.
  4. [Sec. 4.3.3] The text says 'Table 13 reports the RBO values,' but the table reports Robustness Scores (Consistency × Responsiveness), not RBO values. Please correct the text and caption.
  5. [Figure 1] Figure 1 is referenced in the introduction, but the Procrustes distance is not defined in the main text, and the axes are unlabeled. This makes the figure hard to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed metric is an explicitly defined framework, and the empirical rankings are applications of that definition rather than predictions forced by fitted inputs or self-citations.

full rationale

The paper's central equations (Eq. 2–4) define the Robustness Metric as the product of Consistency (median RBO under class-preserving perturbations) and Responsiveness (AUC separating RBO values by class-change label). These are explicit definitions, not derivations from the target result. The finding that GradCAM++ is most robust is the outcome of applying this definition to measured RBO values; it is not a tautology because the metric could in principle rank a different CAM method depending on the data. Consistency and Responsiveness are different functions of the same underlying RBO data, but neither is constructed to reproduce the other, and the product in Eq. 2 is presented as a stated hypothesis, not as a proved theorem. The paper explicitly acknowledges in Table 1 that it 'does not evaluate ground-truth alignment,' so it does not claim external ground-truth validation; this limits scope but does not create circularity. The in-sample AUC computation in Section 2.5 and Algorithm 1 is a statistical weakness (no held-out evaluation, potential undefined AUC when class changes are rare, and Table 13 contains inconsistent/repeated rows), but it is not a fitted parameter renamed as a prediction: the AUC is the very statistic used to define Responsiveness. The paper contains no load-bearing self-citations; it relies on external methods such as RBO [15], QuickShift [39], and standard CAM implementations. The limitations in Section 5 (e.g., not addressing CT/MRI/X-ray modalities) are explicitly stated and do not reduce the central claim to its own inputs. Overall, no circular step meeting the evidence threshold is present.

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

The central claim rests mainly on domain assumptions about the meaningfulness of superpixel-ranked CAM intensities and about the validity of an in-sample fitted classifier AUC as a measure of responsiveness. There are no invented physical entities, but there are several hand-chosen parameters (RBO p, segmentation settings, noise levels). The paper does not derive the metric from first principles.

free parameters (4)
  • RBO persistence p = 0.9
    Chosen by hand in Section 2.5 (Eq. 3); controls how strongly top-ranked segments dominate the consistency score.
  • QuickShift segmentation parameters = kernel size 10, max dist 200, ratio 0.5
    Chosen by hand in Section 3; affect the partition into superpixels and hence all segment rankings.
  • Noise severity levels = Table 3 values; in Table 13 medium equals high for every entry
    Hand-selected low/medium/high settings. The severity ablation table reports identical medium and high rows, so the chosen levels do not behave as intended or the table is erroneous.
  • Linear classifier for responsiveness = not reported
    A linear classifier is trained to map RBO to class-change label; its parameters are fitted to the evaluation data (Algorithm 1), and AUC is computed on the same data.
assumptions (4)
  • domain assumption Superpixels from QuickShift are semantically coherent regions, so ranking segments by mean CAM intensity yields a meaningful explanation.
    Section 2.1 and 2.4; the entire consistency/responsiveness pipeline depends on this.
  • domain assumption RBO with persistence p=0.9 is an appropriate similarity measure for segment importance rankings.
    Section 2.6; the choice of p is not justified by an external criterion.
  • domain assumption AUC of a linear classifier trained on RBO scores validly quantifies CAM responsiveness to prediction changes.
    Section 2.5; this is the operative definition of responsiveness and is assumed to reflect a property of the CAM rather than of the classifier.
  • domain assumption Pretrained ImageNet models are representative for evaluating CAM robustness on the five chosen datasets, including medical images.
    Section 3; models are taken from the PyTorch library without re-training on per-dataset distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing the Noise Robustness of Class Activation Maps: A Framework for Reliable Model Interpretability." pith.science (2026). https://pith.science/paper/Y6B47KDP

@misc{pith2026250818154,
  author       = {Pith},
  title        = {Pith review of: Assessing the Noise Robustness of Class Activation Maps: A Framework for Reliable Model Interpretability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y6B47KDP}},
  note         = {Machine review of arXiv:2508.18154}
}
read the original abstract

Class Activation Maps (CAMs) are one of the important methods for visualizing regions used by deep learning models. Yet their robustness to different noise remains underexplored. In this work, we evaluate and report the resilience of various CAM methods for different noise perturbations across multiple architectures and datasets. By analyzing the influence of different noise types on CAM explanations, we assess the susceptibility to noise and the extent to which dataset characteristics may impact explanation stability. The findings highlight considerable variability in noise sensitivity for various CAMs. We propose a robustness metric for CAMs that captures two key properties: consistency and responsiveness. Consistency reflects the ability of CAMs to remain stable under input perturbations that do not alter the predicted class, while responsiveness measures the sensitivity of CAMs to changes in the prediction caused by such perturbations. The metric is evaluated empirically across models, different perturbations, and datasets along with complementary statistical tests to exemplify the applicability of our proposed approach.

Figures

Figures reproduced from arXiv: 2508.18154 by the authors.

Figure 1
Figure 1. Procrustes distance computed for different networks using Gaussian noise. The weight space in plots shows interesting trends with fluctuations in Procrustes distance distribution when evaluated with multiple noise values. weight space across different CAM methods, where the Procrustes distance 1 [11] between the activation values for original and noisy images differs across various models. Robustness is therefore cr… view at source ↗
Figure 2
Figure 2. Comparison of Stability Ratio (l1) [14] and RBO values. The reference image is without perturbation, whereas others are perturbed. The RBO metric tends to perform relatively stable in both cases and the Stability Ratio, i.e., normalized l1[14] deviates significantly. S(x, x′ , ex, ex′) = max x′ ∥ex − ex′∥ ∥x − x ′∥ , ∀x ′ ∈ Nx; ˆyx = ˆyx′ (1) where, ∥ex − ex′∥ is the l1 distance between explanations generated from t… view at source ↗
Figure 3
Figure 3. Proposed framework for CAM robustness evaluation. The process includes image seg￾mentation, perturbation, CAM computation, saliency ranking, and robustness evaluation via con￾sistency and responsiveness. 2.1. Image Segmentation Each image Ik is first segmented into m visually coherent regions, producing a segmentation map Sk = {s1, s2, . . . , sm}. We use QuickShift [39] as the seg￾mentation algorithm, and the resul… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Distribution of Robustness Scores for different CAM methods across eight noise pertur￾bations on the ImageNet dataset using ResNet-50. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Distribution of Robustness Scores for different CAM methods across eight noise pertur￾bations on the ImageNet dataset using VGG-19. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Distribution of Robustness Scores for different CAM methods across eight noise pertur￾bations on the ImageNet dataset using InceptionV3. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Distribution of Robustness Scores for different CAM methods across eight noise pertur￾bations on the ImageNet dataset using Vision Transformer (ViT). 20 [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of CAMs under different perturbations [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]
Figure 9
Figure 9. Figure 9: Comparison under class-preserving vs. class-changing noise [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 48 canonical work pages

  1. [1]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,

    C. Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,” Nature Machine In- telligence, vol. 1, no. 5, pp. 206–215, 2019

    work page 2019

  2. [2]

    The role of explainability in creating trustworthy artificial intelligence for health care: A comprehen- sive survey of the terminology, design choices, and evaluation strategies,

    A. F. Markus, J. A. Kors, and P. R. Rijnbeek, “The role of explainability in creating trustworthy artificial intelligence for health care: A comprehen- sive survey of the terminology, design choices, and evaluation strategies,” Journal of Biomedical Informatics, vol. 113, p. 103655, 2021

    work page 2021

  3. [3]

    Grad-cam++: Improved visual explanations for deep convolutional net- works,

    A. Chattopadhyay, A. Sarkar, P. Howlader, and V . N. Balasubramanian, “Grad-cam++: Improved visual explanations for deep convolutional net- works,” pp. 839–847, 2018

    work page 2018

  4. [4]

    Axiom-based grad-cam: Towards accu- rate visual explanations for cnns,

    R. Fu, X. Li, Y . Lu, and X. Wang, “Axiom-based grad-cam: Towards accu- rate visual explanations for cnns,” in CVPR, pp. 168–177, 2020

    work page 2020

  5. [5]

    Ablation-cam: Visual expla- nations for deep convolutional networks via gradient-free localization,

    S. Desai, K. Kolar, M. Sadeghi, and S. Yeung, “Ablation-cam: Visual expla- nations for deep convolutional networks via gradient-free localization,” in CVPRW, pp. 973–982, 2020

    work page 2020

  6. [6]

    Hirescam: High-resolution class activa- tion mapping,

    O. Dovrat, I. Mosseri, and A. Tal, “Hirescam: High-resolution class activa- tion mapping,” in ICCV, pp. 213–223, 2019

    work page 2019

  7. [7]

    Eigen-cam: Class activation mapping using prin- cipal components,

    Muhammad and authors, “Eigen-cam: Class activation mapping using prin- cipal components,” IEEE Sign. Process. Letters , vol. 27, no. 8, pp. 1344– 1348, 2020. 36

    work page 2020

  8. [8]

    Revisiting the evaluation of class activation mapping for explainability: A novel metric and experi- mental analysis,

    S. Poppi, M. Cornia, L. Baraldi, and R. Cucchiara, “Revisiting the evaluation of class activation mapping for explainability: A novel metric and experi- mental analysis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2299–2304, 2021

    work page 2021

Show all 55 references
  1. [9]

    Towards trustable explainable ai,

    A. Ignatiev, “Towards trustable explainable ai,” in Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pp. 5154–5158, 7 2020

    2020

  2. [10]

    Rethinking positive aggrega- tion and propagation of gradients in gradient-based saliency methods,

    A. Khakzar, S. Baselizadeh, and N. Navab, “Rethinking positive aggrega- tion and propagation of gradients in gradient-based saliency methods,”arXiv preprint arXiv:2012.00362, 2020

    2012 arXiv

  3. [11]

    Procrustes-based dis- tances for exploring between-matrices similarity,

    A. Andreella, R. De Santis, A. Vesely, and L. Finos, “Procrustes-based dis- tances for exploring between-matrices similarity,”Statistical Methods & Ap- plications, vol. 32, no. 3, pp. 867–882, 2023

    2023

  4. [12]

    Framework for evaluating faithfulness of local explanations,

    S. Dasgupta, N. Frost, and M. Moshkovitz, “Framework for evaluating faithfulness of local explanations,” in International Conference on Machine Learning, pp. 4794–4815, PMLR, 2022

    2022

  5. [13]

    On the robustness of interpretability methods,

    D. Alvarez-Melis and T. Jaakkola, “On the robustness of interpretability methods,” in Proceedings of the 32nd Conference on Neural Information Processing Systems (NeurIPS), pp. 9172–9183, 2018

    2018

  6. [14]

    Rethinking stability for attribution-based explanations,

    C. Agarwal, N. Johnson, M. Pawelczyk, S. Krishna, E. Saxena, M. Zitnik, and H. Lakkaraju, “Rethinking stability for attribution-based explanations,” arXiv preprint arXiv:2203.06877, 2022. 37

    2022 arXiv

  7. [15]

    A similarity measure for indefinite rankings,

    W. Webber, A. Moffat, and J. Zobel, “A similarity measure for indefinite rankings,” ACM Transactions on Information Systems (TOIS), vol. 28, no. 4, pp. 1–38, 2010

    2010

  8. [16]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Ba- tra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on com- puter vision (ICCV), pp. 618–626, 2017

    2017

  9. [17]

    Ex- plaining nonlinear classification decisions with deep taylor decomposition,

    G. Montavon, S. Lapuschkin, A. Binder, W. Samek, and K.-R. M ¨uller, “Ex- plaining nonlinear classification decisions with deep taylor decomposition,” in Pattern Recognition, pp. 87–106, Springer, 2017

    2017

  10. [18]

    Pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. M ¨uller, and W. Samek, “Pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 654–662, 2015

    2015

  11. [19]

    Evaluating the visualization of what a deep neural network has learned,

    W. Samek, T. Wiegand, and K.-R. M ¨uller, “Evaluating the visualization of what a deep neural network has learned,”IEEE Transactions on Neural Net- works and Learning Systems, vol. 28, no. 11, pp. 2660–2673, 2017

    2017

  12. [20]

    What is relevant in a text document?: Learning relevance in document representa- tions,

    L. Arras, F. Horn, G. Montavon, K.-R. M ¨uller, and W. Samek, “What is relevant in a text document?: Learning relevance in document representa- tions,” in Conference on Computer Vision and Pattern Recognition Work- shops, pp. 1695–1704, 2017

    2017

  13. [21]

    On the (in)fidelity and sensitivity of explanations,

    C.-K. Yeh, C.-Y . Hsieh, A. Suggala, D. I. Inouye, and P. K. Ravikumar, 38 “On the (in)fidelity and sensitivity of explanations,” in Advances in Neural Information Processing Systems, vol. 32, 2019

    2019

  14. [22]

    Sanity checks for saliency maps,

    J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, “Sanity checks for saliency maps,” in Advances in Neural Information Pro- cessing Systems (NeurIPS), vol. 31, 2018

    2018

  15. [23]

    How good is your expla- nation? algorithmic stability measures to assess the quality of explanations for deep neural networks,

    T. Fel, D. Vigouroux, R. Cad `ene, and T. Serre, “How good is your expla- nation? algorithmic stability measures to assess the quality of explanations for deep neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10880–10889, 2021

    2021

  16. [24]

    A benchmark for in- terpretability methods in deep neural networks,

    S. Hooker, D. Erhan, P.-J. Kindermans, and B. Kim, “A benchmark for in- terpretability methods in deep neural networks,” in Advances in Neural In- formation Processing Systems (NeurIPS), vol. 32, 2019

    2019

  17. [25]

    Ex- plaining deep neural networks and beyond: A review of methods and appli- cations,

    W. Samek, G. Montavon, S. Lapuschkin, C. Anders, and K.-R. M ¨uller, “Ex- plaining deep neural networks and beyond: A review of methods and appli- cations,” Proceedings of the IEEE, vol. 109, no. 3, pp. 247–278, 2021

    2021

  18. [26]

    Sam: the sensitivity of attribution methods to hyperparameters. in 2020 ieee,

    N. Bansal, C. Agarwal, and A. Nguyen, “Sam: the sensitivity of attribution methods to hyperparameters. in 2020 ieee,” inCVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 11–21, 2020

    2020

  19. [27]

    How explainable are adversarially-robust cnns?,

    M. Nourelahi, L. Kotthoff, P. Chen, and A. Nguyen, “How explainable are adversarially-robust cnns?,” arXiv preprint arXiv:2205.13042, 2022

    2022 arXiv

  20. [28]

    Sanity checks for saliency maps,

    J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, “Sanity checks for saliency maps,” in Advances in Neural Information 39 Processing Systems (S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, eds.), vol. 31, Curran Associat...

    2018

  21. [29]

    On the (in)fidelity and sensitivity of explanations,

    C.-K. Yeh, C.-J. Hsieh, A. Suggala, D. Inouye, and P. Ravikumar, “On the (in)fidelity and sensitivity of explanations,” in Advances in Neural Informa- tion Processing Systems (NeurIPS), vol. 32, 2019

    2019

  22. [30]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conference on Computer Vision and Pattern Recogni- tion, pp. 770–778, 2016

    2016

  23. [31]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint, vol. arXiv:1409.1556, 2014

    2014 arXiv

  24. [32]

    Rethinking the inception architecture for computer vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” in IEEE Conference on Com- puter Vision and Pattern Recognition, pp. 2818–2826, 2016

    2016

  25. [33]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learni...

    2021 arXiv

  26. [34]

    Asirra: A captcha that ex- ploits interest-aligned manual image categorization,

    J. Elson, J. J. Douceur, J. Howell, and J. Saul, “Asirra: A captcha that ex- ploits interest-aligned manual image categorization,” in Proceedings of 14th ACM Conference on Computer and Communications Security (CCS), Asso- ciation for Computing Machinery, Inc., October 2007. 40

    2007

  27. [35]

    Ima- geNet Large Scale Visual Recognition Challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei, “Ima- geNet Large Scale Visual Recognition Challenge,” International Journal of Computer Vision (IJCV), vol. 115, no. 3, pp. 211–252, 2015

    2015

  28. [36]

    Cats and dogs,

    O. M. Parkhi, A. Vedaldi, A. Zisserman, and C. V . Jawahar, “Cats and dogs,” in IEEE Conference on Computer Vision and Pattern Recognition, 2012

    2012

  29. [37]

    The ham10000 dataset: A large collection of multi-source dermatoscopic images of common pigmented skin lesions,

    P. Tschandl, C. Rosendahl, and H. Kittler, “The ham10000 dataset: A large collection of multi-source dermatoscopic images of common pigmented skin lesions,” Scientific data, vol. 5, no. 1, pp. 1–9, 2018

    2018

  30. [38]

    Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,

    L. Fei-Fei, R. Fergus, and P. Perona, “Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,” in 2004 Conference on Computer Vision and Pattern Recognition Workshop, pp. 178–178, IEEE, 2004

    2004

  31. [39]

    Quick shift and kernel methods for mode seek- ing,

    A. Vedaldi and S. Soatto, “Quick shift and kernel methods for mode seek- ing,” in European Conference on Computer Vision, pp. 705–718, Springer, 2008

    2008

  32. [40]

    A similarity measure for indefinite rankings,

    W. Webber, A. Moffat, and J. Zobel, “A similarity measure for indefinite rankings,” ACM Trans. Inf. Syst., vol. 28, Nov. 2010

    2010

  33. [41]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learni...

    2021

  34. [42]

    Icev2: Interpretability, comprehensiveness, and explainability in vision transformer,

    H. Choi, S. Jin, and K. Han, “Icev2: Interpretability, comprehensiveness, and explainability in vision transformer,”International Journal of Computer Vision, pp. 1–18, 2024

    2024

  35. [43]

    JPEG Compression Standard,

    W. Kou, “JPEG Compression Standard,” in Digital Image Compression , vol. 333 of The Springer International Series in Engineering and Computer Science, Springer, Boston, MA, 1995

    1995

  36. [44]

    Comparison of direct blind deconvolu- tion methods for motion-blurred images,

    Y . Yitzhaky and N. S. Kopeika, “Comparison of direct blind deconvolu- tion methods for motion-blurred images,” Applied Optics, vol. 38, no. 20, pp. 4325–4332, 1999

    1999

  37. [45]

    Explaining and Harnessing Adversarial Examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and Harnessing Adversarial Examples,” arXiv preprint arXiv:1412.6572, 2015

    2015 arXiv

  38. [46]

    Towards Deep Learning Models Resistant to Adversarial Attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks,” arXiv preprint arXiv:1706.06083, 2018

    2018 arXiv

  39. [47]

    Towards evaluating the robustness of neural net- works,

    N. Carlini and D. Wagner, “Towards evaluating the robustness of neural net- works,” 2017 IEEE Symposium on Security and Privacy (SP) , pp. 39–57, 2017

    2017

  40. [48]

    Slic superpixels compared to state-of-the-art superpixel methods,

    R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. S ¨usstrunk, “Slic superpixels compared to state-of-the-art superpixel methods,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 11, pp. 2274–2282, 2012

    2012

  41. [49]

    Efficient graph-based image 42 segmentation,

    P. F. Felzenszwalb and D. P. Huttenlocher, “Efficient graph-based image 42 segmentation,” International Journal of Computer Vision , vol. 59, no. 2, pp. 167–181, 2004

    2004

  42. [50]

    A new measure of rank correlation,

    M. G. Kendall, “A new measure of rank correlation,” Biometrika, vol. 30, no. 1/2, pp. 81–93, 1938

    1938

  43. [51]

    The proof and measurement of association between two things,

    C. Spearman, “The proof and measurement of association between two things,” The American Journal of Psychology , vol. 15, no. 1, pp. 72–101, 1904

    1904

  44. [52]

    The problem of m rankings,

    M. G. Kendall and B. B. Smith, “The problem of m rankings,” The Annals of Mathematical Statistics, vol. 10, no. 3, pp. 275–287, 1939

    1939

  45. [53]

    Smoothgrad: removing noise by adding noise,

    D. Smilkov, N. Thorat, B. Kim, F. Vi´egas, and M. Wattenberg, “Smoothgrad: removing noise by adding noise,” in Workshop on Visualization for Deep Learning, ICML, 2017

    2017

  46. [54]

    Probabilistic pixel attribution: Interpreting deep models with statistical inference,

    H. Hegde, S. Kejriwal, and J. Pujara, “Probabilistic pixel attribution: Interpreting deep models with statistical inference,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10899–10908, 2022

    2022

  47. [55]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” inInternational Conference on Learning Representations (ICLR), 2018. 43 Appendix A. Appendix Section Appendix A.1. Why is the RBO Better than l1 Norm fo...

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.