Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

ConvAD replaces input occlusion in post-hoc explanation with neuron deactivation in a CNN forward pass, yielding more robust causal explanations with no retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 12:56 UTC pith:V3KOROWK

load-bearing objection The AD forward pass is a genuinely new idea and the equivalence theorem looks right, but the headline 30-40% robustness gain isn't clean because AD explanations are larger and robustness isn't size-controlled. the 3 major comments →

arxiv 2510.01038 v2 pith:V3KOROWK submitted 2025-10-01 cs.AI cs.CVcs.LG

Activation-Deactivation: A General Framework for Robust Post-hoc Explainable AI

Akchunya Chanchal , David A. Kelly , Hana Chockler This is my paper
classification cs.AI cs.CVcs.LG
keywords convadperturbationactivation-deactivationexplanationsinputmethodsqualityrobust
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

When an image classifier is explained, common methods occlude (mask) parts of the input and see how the prediction changes. But the masked image is out-of-distribution, and the choice of mask fill value (black, gray, blur, noise) requires domain knowledge. The authors propose a different route: keep the input intact, but deactivate the internal neurons that correspond to the masked region. This is done by propagating a binary mask through the network, tracking which intermediate features are affected by the masked input region, and zeroing them at checkpoints (after convolutions or dimension-changing operations). Their algorithm, ConvAD, works on trained CNNs without fine-tuning. They prove that with an all-ones mask (no occlusions), the ConvAD network output is identical to the original network output. The explanation is still computed by their rex tool, but the 'mutant' images the rex tool uses are replaced by internally deactivated versions of the same image. They evaluate on ResNet-50, RegNetY-12GF, and EfficientNet-V2 across ImageNet-1k, ImageNet-v2, CalTech-256, and PASCAL-VOC. Robustness of explanations is measured by pasting the explanation onto solid-color and IID backgrounds (N=150) and checking classification accuracy. They report AD explanations are consistently more robust than occlusion-based explanations with Min/Max/Avg/Zero mask values, with gains of 30-40% (up to 62.5%) at various confidence thresholds.

Core claim

Theorem 1: Given a model N and the result N' of applying ConvAD to N, the output of N' is equal to the output of N on all inputs without occlusions (that is, without an accompanying binary mask). Additionally, the central empirical claim: AD explanations are significantly more robust than occlusion-based explanations — 'AD consistently outperforms the best masking values by 30-40% across all thresholds' (Section 5).

Load-bearing premise

Assumption 1 in Appendix C: 'The dimensionality reductions/expansions are due to parametric operations or external additive/subtractive procedures to the intermediate representations.' ConvAD's checkpoint placement and the proof of Theorem 1 rely on this. If a CNN uses input-dependent or data-dependent shape changes that are not parametric external procedures (e.g., adaptive pooling where output size depends on input spatial size, or RNN-style variable-length dynamics), the mask propagation breaks. Also, for real CNNs, padding at convolutions is treated as 'external additive' and its masking is 'an arbitrary decision' (Appendix C, external additive case); the mask propagation for padding is not a faithful tracking of input influence, which can bias explanations near borders.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces Activation-Deactivation (AD), a forward-pass paradigm for post-hoc explainability of image classifiers that avoids input occlusion (and thus out-of-distribution mutants) by instead deactivating internal activations corresponding to masked input regions. The authors present ConvAD, a drop-in mechanism for CNNs, and prove (Theorem 1, Appendix C) that ConvAD leaves the network output unchanged on unmasked inputs under an assumption about dimensionality-altering operations. They evaluate ConvAD on three CNN architectures and four datasets, comparing AD explanations against occlusion-based baselines (Min, Max, Avg, Zero) using a robustness proxy: planting the explanation's pixel set on 100 solid-color or IID backgrounds and measuring how often the original class is recovered. The paper reports that AD consistently outperforms the best masking strategy by 30-40% across confidence thresholds, and that AD explanations are generally larger than the baselines.

Significance. If the empirical claims survive scrutiny, the AD framework is a meaningful step toward robust post-hoc explainability: it removes the need to choose an occlusion value, avoids out-of-distribution perturbations, and is applicable to trained CNNs without fine-tuning. The paper also provides a theoretical connection to causal explanations via restricted depth-2 causal models. However, the central empirical claim (30-40% robustness improvement) is potentially confounded by explanation size, and the proof of Theorem 1 is conditional on restrictive assumptions. The paper ships code and detailed supplementary results, which is a strength, but the evaluation lacks statistical grounding (no error bars, no significance tests).

major comments (3)
  1. [Section 5.3 and Definition 2] The headline robustness claim is confounded by explanation size. The robustness metric (Def. 2) measures whether the model recovers the original class after planting the entire explanation pixel set onto a background. A larger pixel set contains more of the original image and mechanically raises this probability. Table 1 shows AD explanations are consistently larger than the best masking baseline, from +0.9% to +19.53% depending on model/dataset/threshold. The paper notes this but never matches explanation sizes when computing ρ-robustness. To support the claimed 30-40% improvement, the authors should either (a) subsample larger AD explanations to the size of the baseline before planting, (b) report a size-normalized robustness measure, or (c) explicitly demonstrate that the robustness gain persists when size is controlled. Without this, the central empirical claim may be an artifact of
  2. [Appendix C, Assumption 1 and Theorem 1] The proof of Theorem 1 is explicitly conditional on Assumption 1: 'The dimensionality reductions/expansions are due to parametric operations or external additive/subtractive procedures to the intermediate representations.' The authors further state that for external additive effects (including padding), masking or unmasking is 'an arbitrary decision.' This means Theorem 1 does not cover standard CNNs with padding, nor does it cover adaptive pooling or input-dependent shape changes. Since padding is present in essentially all CNNs, the claim in Section 4 that ConvAD is a general drop-in mechanism is stronger than what is proved. The main text should state the scope of Theorem 1 and clarify the padding policy used in the experiments; the arbitrary treatment of padding may introduce border biases in explanations.
  3. [Section 5 (Figs. 5, A1) and Table 1] The empirical results are presented as point estimates over a sample of N=150 images, with no error bars, confidence intervals, or significance tests. The paper makes strong comparative claims ('consistently outperforms', '30-40% improvement') that would be much more convincing with per-image paired comparisons and a measure of variance. This is especially important because the size confound (see first major comment) can interact with variability across images and datasets.
minor comments (6)
  1. [Abstract and Section 1] The abstract reports 'up to 62.5%' improvement, while Section 1 states '30-40% across all thresholds.' These numbers should be reconciled, and the exact computation of the improvement should be defined.
  2. [Definition 4] The text says Φ_i returns 'a ratio of masked to unmasked input features,' but the formula computes the fraction of unmasked features (sum of 1(M[k,l]>0) divided by |pos|). The wording should be corrected to 'fraction of unmasked features.'
  3. [Definition 5 and Algorithm 1] In Definition 5, the mask M is described as M∈R^2, which should be R^{m×n} (or the input dimensions). In Algorithm 1, the symbol M is used for both the mask and the model in Theorem 1; this overloading is confusing and should be disambiguated.
  4. [Section 5.1] The statement that CalTech-256 results are 'more mixed at the lower confidence threshold' and 'due to the increased amount of noise in explanations at lower confidence levels' is not supported by any measure of noise. Consider providing a brief explanation or a reference.
  5. [Section 4 and Figure 4] The description of leakage scenarios is clear, but the paper does not say how the ‘random perturbations’ scenario in Fig. 4c is handled in the experiments, given that rex uses partial occlusions. Clarify whether such perturbations actually occur in the evaluation.
  6. [Section 6 / References] The paper relies heavily on unpublished preprints [25] and [26] for conceptual claims (e.g., prime-implicant connection, small-explanation problem). These should be marked as such, or their key results should be summarized to make the paper self-contained.

Circularity Check

2 steps flagged

Minor self-definition and self-citation; central AD derivation and robustness evaluation are not circular.

specific steps
  1. self definitional [Section 3, Lemma 3.1]
    "AD-explanations (Definition 7) are causal explanations (Definition 1) with respect to the set of all possible contexts. The proof follows from the observation that the set of all possible restrictions of the causal model includes the original model."

    Definition 7 already defines an AD-explanation as a causal explanation holding for all restrictions M_N|V' of M_N. Since the original model is the restriction with V'=V, Lemma 3.1 restates the definition rather than proving a new implication. The statement is true but carries no independent content beyond the definition.

  2. self citation load bearing [Section 3, Corollary 3.2]
    "As [26] proved that causal explanations over the set of all possible contexts are prime-implicant explanations, we have that AD-explanations are also prime-implicant explanations (but the inverse does not necessarily hold)."

    Reference [26] is Kelly and Chockler's own arXiv preprint. The corollary's only support for the prime-implicant property is this self-citation, with no independent proof or external check provided in the present paper. This is a minor theoretical remark and is not load-bearing for the paper's central robustness experiments.

full rationale

The central derivation is not circular. ConvAD is defined by an explicit algorithm; Theorem 1 is proved from all-ones mask propagation plus the stated architectural assumption, and the empirical robustness comparison uses an external planting benchmark with rex-generated explanations and fixed hyperparameters (tau=0) rather than fitting any parameter to the reported metric. The main caveat is that the headline 30-40% robustness advantage may be confounded by explanation size (Table 1 shows AD explanations are up to 19.53% larger), but this is an internal-validity concern rather than a definitional reduction: the robustness numbers are not forced by the construction of the AD forward pass. The only genuine reductions I found are formal: Lemma 3.1 is a restatement of Definition 7, and Corollary 3.2 imports the prime-implicant claim from the authors' own prior work. These do not affect the main empirical or algorithmic claims.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The paper's central claim (AD explanations are more robust) depends on the causal model framework from prior work (Chockler-Halpern, rex), the variadic function assumption, and the parametric shape-change assumption. The method's own parameters (τ) are fixed, but the algorithm introduces no new fitted parameters. The 'invented entities' are formal definitions, not physical postulates.

free parameters (2)
  • tau (τ) threshold for position-attribution = 0 for all experiments
    The threshold below which activations are discarded. The paper sets it to 0, avoiding tuning, but the framework allows values >0. It is a free parameter of the method, though not fitted to data.
  • confidence threshold gamma (γ) = 0, 0.1, 0.3, 0.5, 0.7, 0.9
    Used to define explanation extraction level. This is an experimental parameter swept over, not fitted.
axioms (4)
  • domain assumption The neural network computes a variadic function, allowing restriction to a subset of inputs (Section 3, Definition 6-7).
    Restriction of a function to a subset of parameters is not generally defined; the paper assumes a DNN can be treated as variadic. This is needed to connect causal explanations to AD.
  • domain assumption Causal independence between input features (Appendix B).
    Needed for the depth-2 causal model to give AC2 with W empty. Reasonable for images but an assumption.
  • domain assumption Assumption 1 (Appendix C): dimensionality changes are due to parametric operations or external additive/subtractive procedures.
    Needed for ConvAD checkpoint coverage and for the proof of Theorem 1. Excludes pooling/upsampling that depend on input content in non-parametric ways, though adaptive pooling is often parametric.
  • domain assumption Context set K for explanations is defined over permutations/occlusions of the original image, and rex's approximations satisfy EXIC1/EXIC3 but not EXIC2 (Section 5).
    The paper relies on rex's approximate explanations as the base; if rex's approximations are flawed or do not capture the true causal explanation, AD's improvement over occlusion values inherits that.
invented entities (2)
  • Restricted depth-2 causal model (Definition 6) no independent evidence
    purpose: To formally capture AD operation as a restriction of the original causal model.
    This is a theoretical construct introduced in the paper. It has no falsifiable handle outside the paper's own framework.
  • AD-explanation (Definition 7) no independent evidence
    purpose: A new type of explanation defined via all restrictions of the causal model.
    Defined in the paper; claims to be a stronger notion than causal explanation. No external benchmark provided.

pith-pipeline@v1.3.0-alltime-deepseek · 17247 in / 3758 out tokens · 30125 ms · 2026-08-04T12:56:18.832612+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Activation-Deactivation: A General Framework for Robust Post-hoc Explainable AI." pith.science (2026). https://pith.science/paper/V3KOROWK

@misc{pith2026251001038,
  author       = {Pith},
  title        = {Pith review of: Activation-Deactivation: A General Framework for Robust Post-hoc Explainable AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3KOROWK}},
  note         = {Machine review of arXiv:2510.01038}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Perturbation-based explainability methods face criticism due to their reliance on out-of-distribution mutants. This raises doubts about the quality of the explanations. In this paper, we introduce a novel forward pass paradigm, Activation-Deactivation (AD), which obviates the need for perturbation of the input. AD replaces perturbation of input features with switching off parts of the model corresponding to to the intended perturbations. We implement ConvAD, an AD approximation algorithm for CNNs. ConvAD is a drop-in mechanism that can be easily added to any trained CNN and, without any additional training, generates more robust and more transferable explanations. We provide evaluation results across multiple architectures, datasets, methods and perturbation strategies, demonstrating the superior quality of ConvAD compared to the SOTA.

Figures

Figures reproduced from arXiv: 2510.01038 by Akchunya Chanchal, David A. Kelly, Hana Chockler.

Figure 1
Figure 1. Figure 1: An image of an ibex (a) and its explanations using our method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ① The state-of-the-art approach to generating perturbations for post hoc explainability: masking parts of the input image. ② Our approach Activation-Deactivation (AD): occluding parts of the model. AD preserves the spatial locality of the unmasked features while removing any consideration of the masked features, remaining in distribution for the input features to the model. We implemented our framework and… view at source ↗
Figure 3
Figure 3. Figure 3: Visual demonstration of applying the position attribution function, in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scenarios where leaks can occur when applying [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: ρ-robustness (rows, from 0 to 1 in steps of 0.2) of explanation against non-informative background on ImageNet-1k, ImageNet-1k v2, CalTech(CT)- 256 and PASCAL-VOC for our models, with different confidence thresholds γ. AD explanations are consistently more robust than ones computed using masking values, for all masking values and with all confidence thresholds. The sets of models and datasets were selected… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. If It's Good Enough for You, It's Good Enough for Me: Transferability of Audio Sufficiencies across Models

    cs.SD 2026-04 unverdicted novelty 7.0

    Transferability analysis finds that minimal sufficient signals transfer across audio models at rates varying by task, around 26% for music genre classification, with some deepfake models showing distinct behaviors not...

Reference graph

Works this paper leans on

52 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    On pixel-wise ex- planations for non-linear classifier decisions by layer-wise relevance propa- gation.PLoS ONE, 10(7), 2015

    Sebastian Bach, Alexander Binder, Gr´ egoire Montavon, Frederick Klauschen, Klaus-Robert M¨ uller, and Wojciech Samek. On pixel-wise ex- planations for non-linear classifier decisions by layer-wise relevance propa- gation.PLoS ONE, 10(7), 2015

    2015

  2. [2]

    S. Beckers. Causal sufficiency and actual causation.Journal of Philosoph- ical Logic, 50:1341–1374, 2021

    2021

  3. [3]

    Specrex: Explainable AI for Raman spectroscopy.arXiv preprint arXiv:2503.14567, 2025

    Nathan Blake, David A Kelly, Akchunya Chanchal, Sarah Kapllani-Mucaj, Geraint Thomas, and Hana Chockler. Specrex: Explainable AI for Raman spectroscopy.arXiv preprint arXiv:2503.14567, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Melime: Meaningful local explanation for machine learning models

    Tiago Botari, Frederik Hvilshøj, Rafael Izbicki, and Andre CPLF de Car- valho. Melime: Meaningful local explanation for machine learning models. arXiv preprint arXiv:2009.05818, 2020

    Pith/arXiv arXiv 2009

  5. [5]

    Urszula Chajewska and Joseph Y. Halpern. Defining explanation in prob- abilistic systems. InUncertainty in Artificial Intelligence (UAI), pages 62–71. Morgan Kaufmann, 1997

    1997

  6. [6]

    Grad-cam++: Generalized gradient-based visual expla- nations for deep convolutional networks

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual expla- nations for deep convolutional networks. In2018 IEEE winter conference on applications of computer vision (WACV), pages 839–847. IEEE, 2018

    2018

  7. [7]

    Hana Chockler and Joseph Y. Halpern. Responsibility and blame: A structural-model approach.J. Artif. Intell. Res., 22:93–115, 2004

    2004

  8. [8]

    Hana Chockler and Joseph Y. Halpern. Explaining image classifiers, 2024

    2024

  9. [9]

    Causal explanations for image classifiers.arXiv preprint arXiv:2411.08875, 2024

    Hana Chockler, David A Kelly, Daniel Kroening, and Youcheng Sun. Causal explanations for image classifiers.arXiv preprint arXiv:2411.08875, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Ima- genet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Ima- genet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  11. [11]

    Anders, Mar- cel Ackermann, Klaus-Robert M¨ uller, and Pan Kessel

    Ann-Kathrin Dombrowski, Maximilian Alber, Christopher J. Anders, Mar- cel Ackermann, Klaus-Robert M¨ uller, and Pan Kessel. Explanations can be manipulated and geometry is to blame, September 2019. arXiv:1906.07983 [cs, stat]

    Pith/arXiv arXiv 2019

  12. [12]

    Williams, John Winn, and Andrew Zisserman

    Mark Everingham, Luc Gool, Christopher K. Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge.Int. J. Comput. Vision, 88(2):303–338, June 2010

    2010

  13. [13]

    MIT Press, 1988

    Peter G¨ ardenfors.Knowledge in Flux. MIT Press, 1988. 14

    1988

  14. [14]

    Glymour and F

    C. Glymour and F. Wimberly. Actual causes and thought experiments. In J. Campbell, M. O’Rourke, and H. Silverstein, editors,Causation and Explanation, pages 43–67. MIT Press, Cambridge, MA, 2007

    2007

  15. [15]

    Caltech 256, Apr 2022

    Gregory Griffin, Alex Holub, and Pietro Perona. Caltech 256, Apr 2022

    2022

  16. [16]

    N. Hall. Structural equations and causation.Philosophical Studies, 132:109–136, 2007

    2007

  17. [17]

    J. Y. Halpern and J. Pearl. Causes and explanations: a structural- model approach. Part I: causes.British Journal for Philosophy of Science, 56(4):843–887, 2005

    2005

  18. [18]

    Halpern.Actual Causality

    Joseph Y. Halpern.Actual Causality. The MIT Press, 2019

    2019

  19. [19]

    audi- olime: Listenable explanations using source separation.arXiv preprint arXiv:2008.00582, 2020

    Verena Haunschmid, Ethan Manilow, and Gerhard Widmer. audi- olime: Listenable explanations using source separation.arXiv preprint arXiv:2008.00582, 2020

    Pith/arXiv arXiv 2008

  20. [20]

    Deep residual learning for image recognition.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  21. [21]

    Free Press, 1965

    Carl Gustav Hempel.Aspects of Scientific Explanation. Free Press, 1965

    1965

  22. [22]

    Hitchcock

    C. Hitchcock. The intransitivity of causation revealed in equations and graphs.Journal of Philosophy, XCVIII(6):273–299, 2001

    2001

  23. [23]

    Hitchcock

    C. Hitchcock. Prevention, preemption, and the principle of sufficient reason. Philosophical Review, 116:495–532, 2007

    2007

  24. [24]

    A benchmark for interpretability methods in deep neural networks

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. InAd- vances in Neural Information Processing Systems, pages 9734–9745, 2019

    2019

  25. [25]

    I am big, you are little; i am right, you are wrong.arXiv preprint arXiv:2507.23509, 2025

    David A Kelly, Akchunya Chanchal, and Nathan Blake. I am big, you are little; i am right, you are wrong.arXiv preprint arXiv:2507.23509, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Causal identification of sufficient, con- trastive and complete feature sets in image classification.arXiv preprint arXiv:2507.23497, 2025

    David A Kelly and Hana Chockler. Causal identification of sufficient, con- trastive and complete feature sets in image classification.arXiv preprint arXiv:2507.23497, 2025

    arXiv 2025

  27. [27]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. InAdvances in Neural Information Processing Systems (NeurIPS), volume 30, pages 4765–4774, 2017

    2017

  28. [28]

    Segal time series clas- sification—stable explanations using a generative model and an adaptive weighting method for lime.Neural Networks, 176:106345, 2024

    Han Meng, Christian Wagner, and Isaac Triguero. Segal time series clas- sification—stable explanations using a generative model and an adaptive weighting method for lime.Neural Networks, 176:106345, 2024. 15

    2024

  29. [29]

    Local interpretable model-agnostic explanations for music content analysis

    Saumitra Mishra, Bob L Sturm, and Simon Dixon. Local interpretable model-agnostic explanations for music content analysis. InISMIR, vol- ume 53, pages 537–543, 2017

    2017

  30. [30]

    Morgan Kauf- mann, 1988

    Judea Pearl.Probabilistic Reasoning in Intelligent Systems. Morgan Kauf- mann, 1988

    1988

  31. [31]

    RISE: randomized input sam- pling for explanation of black-box models

    Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: randomized input sam- pling for explanation of black-box models. InBritish Machine Vision Con- ference (BMVC). BMV A Press, 2018

    2018

  32. [32]

    Designing network design spaces

    Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Doll´ ar. Designing network design spaces. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428–10436, 2020

    2020

  33. [33]

    Do imagenet classifiers generalize to imagenet? InInternational conference on machine learning, pages 5389–5400

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? InInternational conference on machine learning, pages 5389–5400. PMLR, 2019

    2019

  34. [34]

    Why should I trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. “Why should I trust you?” Explaining the predictions of any classifier. InKnowledge Discovery and Data Mining (KDD), pages 1135–1144. ACM, 2016

    2016

  35. [35]

    Anchors: high-precision model-agnostic explanations

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Anchors: high-precision model-agnostic explanations. InProceedings of the Thirty- Second AAAI Conference on Artificial Intelligence and Thirtieth In- novative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI’18/IAAI’...

    2018

  36. [36]

    Salmon.Four Decades of Scientific Explanation

    Wesley C. Salmon.Four Decades of Scientific Explanation. University of Minnesota Press, 1989

    1989

  37. [37]

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

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE international conference on computer vision, pages 618–626, 2017

    2017

  38. [38]

    Learning im- portant features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning im- portant features through propagating activation differences. InICML, vol- ume 70, pages 3145–3153. JMLR.org, 2017

    2017

  39. [39]

    Deep inside con- volutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside con- volutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013. 16

    Pith/arXiv arXiv 2013

  40. [40]

    Limesegment: Meaningful, realistic time series explanations

    Torty Sivill and Peter Flach. Limesegment: Meaningful, realistic time series explanations. InInternational Conference on Artificial Intelligence and Statistics, pages 3418–3433. PMLR, 2022

    2022

  41. [41]

    Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods

    Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation Methods. InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society, AIES ’20, pages 180–186, New York, NY, USA, February 2020. Association for Computing Machinery

    2020

  42. [42]

    Riedmiller

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Mar- tin A. Riedmiller. Striving for simplicity: The all convolutional net. In ICLR (Workshop Track), 2015

    2015

  43. [43]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. InInternational Conference on Machine Learning, pages 3319–3328. PMLR, 2017

    2017

  44. [44]

    Efficientnetv2: Smaller models and faster training

    Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. InInternational conference on machine learning, pages 10096– 10106. PMLR, 2021

    2021

  45. [45]

    When can you trust your explanations? a robustness analysis on feature importances.arXiv preprint arXiv:2406.14349, 2024

    Ilaria Vascotto, Alex Rodriguez, Alessandro Bonaita, and Luca Bortolussi. When can you trust your explanations? a robustness analysis on feature importances.arXiv preprint arXiv:2406.14349, 2024

    arXiv 2024

  46. [46]

    B. Weslake. A partial theory of actual causation.British Journal for the Philosophy of Science, 2015. To appear

    2015

  47. [47]

    Woodward.Making Things Happen: A Theory of Causal Explanation

    J. Woodward.Making Things Happen: A Theory of Causal Explanation. Oxford University Press, Oxford, U.K., 2003

    2003

  48. [48]

    Ml-loo: Detecting adversarial examples with feature attribution

    Puyudi Yang, Jianbo Chen, Cho-Jui Hsieh, Jane-Ling Wang, and Michael Jordan. Ml-loo: Detecting adversarial examples with feature attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6639–6647, 2020

    2020

  49. [49]

    Zhichao Yang, Liang Meng, Wenjian Rui, Liang Shen, Sen Zhao, and Lu Shi. Enhancing explainability in raman spectroscopy classification with shap and spectral segmentation.Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 344:126394, 2026

    2026

  50. [50]

    Visualizing and understanding con- volutional networks

    Matthew D Zeiler and Rob Fergus. Visualizing and understanding con- volutional networks. InEuropean conference on computer vision, pages 818–833. Springer, 2014

    2014

  51. [51]

    Baylime: Bayesian local interpretable model-agnostic explanations

    Xingyu Zhao, Wei Huang, Xiaowei Huang, Valentin Robu, and David Flynn. Baylime: Bayesian local interpretable model-agnostic explanations. InUncertainty in artificial intelligence, pages 887–896. PMLR, 2021. 17

    2021

  52. [52]

    F-fidelity: A robust framework for faith- fulness evaluation of explainable ai.arXiv preprint arXiv:2410.02970, 2024

    Xu Zheng, Farhad Shirani, Zhuomin Chen, Chaohao Lin, Wei Cheng, Wenbo Guo, and Dongsheng Luo. F-fidelity: A robust framework for faith- fulness evaluation of explainable ai.arXiv preprint arXiv:2410.02970, 2024. 18 Appendix In this appendix to our work on Activation Deactivation, we provide: (A) Robustness against IID examples. . . . . . . . . . . . . . ....

    Pith/arXiv arXiv 2024