Pith. sign in

REVIEW 5 major objections 6 minor 31 references

CASE: Contrastive Activation for Saliency Estimation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Many saliency maps ignore the class label they are asked to explain; CASE removes shared evidence and restores class-specific explanations.

desk verdict CASE is a simple, clearly written contrastive Grad-CAM variant, but its main evidence for class-specificity rests on a diagnostic that random label-conditioned maps would also pass. read the letter →

arxiv 2506.07327 v3 pith:VWTL475K submitted 2025-06-08 cs.CV cs.LG

Dane Williamson , Yangfeng Ji , Matthew Dwyer This is my paper
classification cs.CVcs.LG
keywords saliencymapsclasssensitivityGrad-CAMcontrastiveexplanationfeatureagreementconfusionmatrixCNNattributionexplainableAI
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

This paper tries to establish that a central failure of popular saliency methods is statistical, not just visual: for the same image, heatmaps produced for different class labels overlap so strongly that the methods often cannot be said to distinguish classes at all. Across four architectures and two datasets, the paper's top-5% feature-agreement diagnostic rejects class sensitivity for most methods on most models, and the authors argue the failure is structural rather than tied to a particular network. To fix it, they propose CASE, which subtracts from the target class's gradient the component shared with the average gradient of the classes the model most often confuses, leaving a residual that highlights what is unique to the predicted class. On the same diagnostic, CASE is the only method that separates top-1 from top-2 explanations on every architecture, and on a confidence-drop fidelity test it matches the stronger baselines on most models.

What carries the argument

The load-bearing object is the orthogonal projection in Equation 5, which removes from the target gradient $\gamma_u$ the part aligned with the mean gradient $\bar{\gamma}$ of the $k$ most confused classes: $\gamma_u^\perp = \gamma_u - \frac{\langle \gamma_u, \bar{\gamma}\rangle}{\|\bar{\gamma}\|^2 + \epsilon}\bar{\gamma}$. The saliency map is then $\mathrm{ReLU}(\gamma_u^\perp \cdot A)$ upsampled to input resolution. The evaluation machinery mirrors this: the diagnostic compares top-$k$ feature agreement $F(E, E', k)$ between class explanations and tests the null that median agreement is at least 50% with a one-sided Wilcoxon signed-rank test, while the confusion matrix computed on a held-out validation set supplies the contrast classes that make the projection behaviorally grounded. Read charitably, the subtraction isolates the set $U_V^u = D_u \setminus \bigcup_{v\in V} D_v$ of features that contribute positively to the target class but to none of the classes the model confuses with it.

What would settle it

On images where the top two predicted classes share an obvious visual component (for example, a shark fin visible in both the top-1 and top-2 explanation), compute CASE's map and then ablate the two halves of its decomposition separately: the projected-away shared direction $\bar{\gamma}\cdot A$ and the residual $\gamma_u^\perp\cdot A$. If removing the shared direction does not lower the confused class's logit, or if the residual's top pixels do not overlap the image region a human or a pixel-level ground truth identifies as distinguishing the two classes, then the orthogonality assumption fails and CASE's class specificity is an artifact of the overlap metric rather than genuine discriminative evidence.

Watch

Extended reading notes

Core claim

The paper's central claim is that class-insensitivity is a measurable, structural property of activation-based saliency methods: Grad-CAM and its popular variants produce nearly identical heatmaps for competing class labels, failing the paper's test that median top-5% agreement between top-1 and top-2 explanations be below 50%. The proposed fix, CASE, computes the gradient of the target class score with respect to the final convolutional activations, averages the gradients over the top-k classes in the model's confusion matrix, and replaces the target gradient by its orthogonal complement with respect to that average. The resulting saliency map therefore emphasizes features not shared with the classes the model itself treats as alternatives. The authors claim this makes explanations more class-specific without sacrificing faithfulness, and support that claim with a Wilcoxon-based diagnostic across ConvNeXt, DenseNet, ResNet, and VGG, together with a perturbation-based fidelity comparison.

Load-bearing premise

CASE's formula removes the part of the target class's gradient that points along the average gradient of the confused classes, and the whole method rests on the premise that this removed direction corresponds to evidence genuinely shared between those classes — if gradient geometry does not track shared semantics, the residual maps can be less faithful while still looking class-specific on the overlap test.

Editorial extensions

If this is right

  • CASE is the only method tested that rejects the class-insensitivity null hypothesis on every architecture, including ResNet and VGG, where Grad-CAM and AblationCAM fail outright.
  • The diagnostic is cheap and model-agnostic, so any team that relies on saliency maps can run it before trusting explanations for auditing or debugging.
  • Attribution layer choice can dominate method choice: DenseNet's sparse final convolutional layer yields class-distinct maps for almost all methods, while its dense normalization layer breaks them, and the effect persists across five training seeds.
  • Class specificity does not come at a uniform fidelity cost: on ConvNeXt, CASE beats AblationCAM and LayerCAM in confidence drop, ties the baselines on DenseNet and ResNet, and trails three baselines on VGG.

Reading between the lines

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

  • If gradient orthogonality really tracks semantic exclusivity, the confusion-matrix contrast set could be replaced by a soft or full-class contrast, making CASE applicable where confusion statistics are unreliable; the paper leaves this extension open.
  • The overlap diagnostic could be run per image rather than as a population test, flagging individual explanations whose top-1 and top-2 maps are near-identical; this would give practitioners a usable warning signal on a single explanation.
  • Because DenseNet's channel sparsity reverses the ranking of all baselines, comparative saliency evaluations that fix one attribution layer per architecture may be measuring layer selectivity more than method quality; controlling layer choice across models would test that confound.
  • The weakest link of the method, that the removed gradient direction is genuinely shared evidence, can be checked directly: if ablating the projected-away direction $\bar{\gamma}\cdot A$ fails to lower the confused classes' logits, CASE's specificity is an artifact of the overlap metric rather than a property of the model's decision.
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

5 major / 6 minor

Summary. The paper proposes a diagnostic test for class sensitivity of saliency methods: for each image, it computes the overlap between the top-k salient pixels of the top-1 and top-2 class saliency maps and uses a one-sided Wilcoxon signed-rank test against a 50% median-overlap null hypothesis (Section 4.2, Eq. 8). Applying this test across four architectures, the authors report that many existing CAM-style methods fail to reject the null, while their proposed method CASE consistently rejects it (Table 1). CASE is a contrastive extension of Grad-CAM that subtracts from the target-class gradient its projection onto the mean gradient of the top-k confused classes (Eqs. 2 and 5). The paper also measures faithfulness by ablating top-k salient pixels and comparing confidence drops (Section 5.3, Table 3). The central claim is that CASE produces 'faithful and more class-specific explanations than existing methods' (abstract).

Significance. If the diagnostic were a valid measure of semantic class-specificity and CASE's contrastive projection indeed isolated uniquely discriminative features, the paper would make a useful contribution to the interpretability literature: it addresses a real and well-documented failure mode of activation-based saliency methods, and it proposes a simple, architecture-agnostic modification that requires no retraining. The paper also provides extensive empirical coverage across four architectures and includes robustness checks for training seeds and layer selection, which are valuable in themselves. However, the significance is tempered by the fact that the central evidence for class-specificity rests on a diagnostic that, as formulated, cannot distinguish class-conditional noise from semantically meaningful class-discriminative attribution. The method's mechanism also directly minimizes the very overlap metric used to evaluate it, making the main quantitative claim partly circular.

major comments (5)
  1. [Section 4.2, Eq. (8)] The proposed RQ1 diagnostic is vacuous as a test of semantic class-specificity. The null hypothesis is that median top-5% agreement between top-1 and top-2 maps is at least 50%, but a method that outputs independent random 5%-pixel masks for each class label has expected agreement 5% and would reject the null with p<0.0001. Thus the p-values in Table 1 establish only that maps differ across labels, not that the differences reflect features that are 'uniquely discriminative' in the sense of Eq. (3). The paper needs a noise control: for example, report the RQ1 result for random class-conditioned masks, or for a method whose class label is scrambled, and show that this control fails to reject the null. Without such a control, the diagnostic cannot distinguish class-specific explanations from label-conditioned noise.
  2. [Section 3.3, Eqs. (2) and (5), and Section 5.1] CASE's low top-1/top-2 overlap is guaranteed by construction. The contrast set V in Eq. (2) is selected from the most confused classes, and for the top-1/top-2 comparison the top-2 class will typically be in V. Eq. (5) then removes from γ_u the projection onto the mean contrast gradient, so the feature-agreement metric of Eq. (7) is directly minimized by the method's design. Consequently, Table 1's demonstration that CASE rejects the null does not provide independent evidence that CASE isolates semantically class-specific evidence; it only confirms that the method does what it is engineered to do. To support the central claim, the paper should evaluate CASE with an independent class-specificity measure that is not an explicit objective of the method, such as a comparison of overlap for semantically related versus unrelated class pairs, or an external pointing-game or segmentation-based evaluation.
  3. [Section 5.3, Table 3] The fidelity results do not support the abstract's claim that CASE produces 'faithful and more class-specific explanations than existing methods.' On DenseNet and ResNet, CASE is statistically indistinguishable from most baselines; on VGG, three baselines (Grad-CAM++, LayerCAM, ScoreCAM) produce significantly higher confidence drops than CASE. The discussion in Section 6.2 acknowledges a tradeoff, but the abstract and Section 7 state that CASE 'outperforms or matches existing methods in class specificity and attribution faithfulness,' which is an overstatement given Table 3. The fidelity experiment should be reported as showing parity or mixed results, and the wording of the abstract and conclusions should be revised accordingly.
  4. [Section 4.4 and Contributions] The paper claims evaluation on two standard datasets, ImageNet and CIFAR-100, but the main results in Tables 1, 2, and 3 report only models pretrained on ImageNet. CIFAR-100 appears only in the DenseNet seed-robustness analysis of Section 5.1.1 and in Figure 4. The contributions section states that the diagnostic and CASE are evaluated 'across four architectures and two datasets,' which is not supported by the reported experiments. Either add CIFAR-100 results for all architectures and methods, or explicitly scope the empirical claims to ImageNet with CIFAR-100 as a supplementary robustness check.
  5. [Section 3.4, Eq. (5)] The central design assumption—that orthogonalizing γ_u against the mean contrast gradient isolates the uniquely discriminative feature set U_V^u of Eq. (3)—is not verified against any independent notion of discriminativeness. In a high-dimensional activation space, the residual γ_u^⊥ can be dominated by directions that are merely orthogonal to the low-dimensional span of a few contrast gradients but are not semantically exclusive to class u. The paper should provide evidence that the residual maps correspond to class-specific evidence, for example by showing that CASE maps align better with class-discriminative image regions in a human-annotated localization benchmark, or by demonstrating that CASE's top-k regions are more predictive of the target class than of a random class in a deletion/insertion test that is not based on the overlap metric.
minor comments (6)
  1. [Figure 3 caption] The caption contains a typo: 'V alley' should be 'Valley'.
  2. [Section 5.3] The sentence 'The null hypothesis states that the mean confidence drop of the method is greater than or equal to that of CASE' is consistent with Eq. (9) but could be rephrased for clarity; the current wording makes the direction of the test difficult to parse.
  3. [Table 1] The paper reports 24 Wilcoxon tests without any multiple-comparison correction; with a p<0.05 threshold, several false positives are expected across the table. The authors should report corrected p-values or at least note the multiple-testing issue.
  4. [Section 4.4] The sentence 'All models are evaluated using publicly available pretrained model without fine-tuning' has a subject-verb agreement error and should read 'pretrained models.'
  5. [Equations (5) and (6)] The notation '↑ β' is used before being defined; please define the upsampling factor β explicitly in the text or in Algorithm 1.
  6. [Algorithm 1] The stability term ε in Eq. (5) is introduced but its value is never specified in the experiments; please report the chosen value for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

CASE's RQ1 win is largely by construction: the overlap-based diagnostic is exactly the quantity Eq. (5) suppresses, while the independent fidelity test does not support greater class-specificity.

  1. self definitional [Section 3.3 (Eq. 2), Section 3.5 (Eq. 5), Section 4.2 (Eq. 8), Table 1]
    "V = arg k max v∈C\{u} M [u, v] (2) ... γ⊥ u = γu − ⟨γu, ¯γ⟩ ∥¯γ∥2 + ϵ · ¯γ (5) ... The null hypothesis is that the median agreement between saliency maps for the top-1 and top-2 class labels is greater than or equal to 50%."

    RQ1's pass criterion is low top-1/top-2 overlap (Eq. 8). CASE's contrast set V is the top-k confused classes (Eq. 2); for an input whose top-2 prediction is one of those confused classes, Eq. 5 removes from the target gradient the component parallel to the mean gradient over V, which includes the top-2 class's gradient. The measured agreement therefore drops as a direct algebraic consequence of the construction, not as an independent empirical finding. The paper's evidence for 'more class-specific explanations' (abstract) is Table 1, where CASE rejects the 50% null; no independent check (e.g., a noise control or verification that the orthogonal residual equals the set-difference U^u_V of Eq. 3) distinguishes CASE's low overlap from label-conditioned random maps.

full rationale

The paper's derivation chain has one load-bearing circular segment. RQ1 (Section 4.2) defines class sensitivity through low top-1/top-2 top-5% pixel agreement (Eqs. 7-8). CASE's algorithm (Sections 3.3-3.5) selects as its contrast set the top-k confused classes from the model's confusion matrix and defines its saliency map by subtracting from the target-class gradient the projection onto the mean gradient of that set (Eq. 5). Since the top-2 predicted class on a correctly classified input is, by construction of the confusion matrix, a frequently confused alternative, it will typically lie in V, and Eq. 5 removes a direction aligned with that class's gradient. Hence CASE's low agreement in Table 1 is an algebraic consequence of the method's construction. The paper calls this 'class-specificity' and the abstract claims 'more class-specific explanations,' but the only supporting evidence is the RQ1 p-values; the independent RQ2 fidelity test shows CASE at best matches baselines and on VGG is significantly worse than three methods. The paper does not rely on self-citation chains: citations to Adebayo, Krishna, Kalibhat, and others are external and appropriate; the confusion-matrix and Grad-CAM machinery are standard. The deeper semantic claim that the orthogonal residual equals the uniquely discriminative feature set U^u_V (Eq. 3) is asserted without proof, and the 50% threshold would also be rejected by label-conditioned random maps. Overall, the central 'more class-specific' claim is partially circular/self-definitional, so score 6; the architecture ablations and the fidelity comparison provide independent content but do not rescue the headline class-specificity claim.

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

The central claim rests on the diagnostic design (top-k overlap threshold) and on the geometric assumption that gradient orthogonality equals feature exclusivity. Several evaluation choices (k, epsilon, thresholds) are free parameters not justified by sensitivity analysis.

free parameters (4)
  • k (number of contrast classes) = 3 (stated in Limitations)
    Algorithm 1 takes k as input; the paper says small values (e.g., k=3) suffice but provides no ablation or sensitivity analysis.
  • epsilon (projection stability term) = unspecified
    Eq. 5 uses +epsilon to avoid division by zero; the value is not reported and affects the projection when ||bar_gamma|| is small.
  • tau (active channel threshold) = 0.001
    Eq. 10 uses tau=0.001 to define active channels in the ablation; this is arbitrary and not justified.
  • top-k overlap fraction and null threshold = top 5%; null median agreement 50%
    RQ1 uses top-5% salient pixels and a null median agreement of 50%; these thresholds are chosen without sensitivity analysis and directly determine which methods 'fail' the diagnostic.
assumptions (3)
  • domain assumption Gradients of class scores with respect to bottleneck activations are meaningful attribution signals.
    Standard in the Grad-CAM literature; invoked in Section 3.2 to define gamma_u.
  • domain assumption The confusion matrix computed on a held-out validation set reliably identifies classes the model confuses with the target.
    Section 3.3 selects contrast set V via Eq. 2; assumes confusion counts are stable and representative.
  • ad hoc to paper Linear orthogonalization in gradient space recovers uniquely discriminative features.
    Section 3.4 equates the set difference U_V^u (Eq. 3) with the projection residual (Eq. 5); this equivalence is asserted, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CASE: Contrastive Activation for Saliency Estimation." pith.science (2026). https://pith.science/paper/VWTL475K

@misc{pith2026250607327,
  author       = {Pith},
  title        = {Pith review of: CASE: Contrastive Activation for Saliency Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWTL475K}},
  note         = {Machine review of arXiv:2506.07327}
}
read the original abstract

Saliency methods are widely used to visualize which input features are deemed relevant to a model's prediction. However, their visual plausibility can obscure critical limitations. In this work, we propose a diagnostic test for class sensitivity: a method's ability to distinguish between competing class labels on the same input. Through extensive experiments, we show that many widely used saliency methods produce nearly identical explanations regardless of the class label, calling into question their reliability. We find that class-insensitive behavior persists across architectures and datasets, suggesting the failure mode is structural rather than model-specific. Motivated by these findings, we introduce CASE, a contrastive explanation method that isolates features uniquely discriminative for the predicted class. We evaluate CASE using the proposed diagnostic and a perturbation-based fidelity test, and show that it produces faithful and more class-specific explanations than existing methods.

Figures

Figures reproduced from arXiv: 2506.07327 by the authors.

Figure 1
Figure 1. Saliency maps for different class labels on the same image using Grad-CAM. Despite [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CASE algorithm. Gradients are backpropagated for the target class (blue) and contrast classes (red, green). The contrast gradients are averaged, and the target gradient is projected onto this average. The projection is then subtracted from the original target gradient to isolate a class-specific signal. The resulting orthogonal component is used to compute the weights for the final saliency map. intr… view at source ↗
Figure 3
Figure 3. Class sensitivity comparison across saliency methods. Each row shows the original image [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Mean agreement scores across seeds for each method (DenseNet, CIFAR-100). Lower [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Kernel density estimate (KDE) of mean active channels per class [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Distribution of per-class mean active channel counts across architectures. DenseNet [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Agreement score distributions for all saliency methods (VGG). Each histogram shows the [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Agreement score distributions for all saliency methods (ResNet). Each histogram shows [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Agreement score distributions for all saliency methods (ConvNeXt). Each histogram [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Agreement score distributions for all saliency methods (DenseNet (Conv)). Each his [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Agreement score distributions for all saliency methods (DenseNet (Norm5)). Each [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 7 canonical work pages

  1. [1]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS'18, page 9525–9536, Red Hook, NY, USA, 2018. Curran Associates Inc

    work page 2018

  2. [2]

    Evaluating saliency map explanations for convolutional neural networks: a user study

    Ahmed Alqaraawi, Martin Schuessler, Philipp Wei , Enrico Costanza, and Nadia Berthouze. Evaluating saliency map explanations for convolutional neural networks: a user study. In Proceedings of the 25th International Conference on Intelligent User Interfaces, IUI '20, page 275–285, New York, NY, USA, 2020. Association for Computing Machinery. ISBN 978145037...

    arXiv 2020

  3. [3]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation

    Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLOS ONE, 10 0 (7): 0 1--46, 07 2015. doi:10.1371/journal.pone.0130140. URL https://doi.org/10.1371/journal.pone.0130140

    doi:10.1371/journal.pone.0130140 2015

  4. [4]

    Network dissection: Quantifying interpretability of deep visual representations

    David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. CoRR, abs/1704.05796, 2017. URL http://arxiv.org/abs/1704.05796

    arXiv 2017

  5. [5]

    Explaining image classifiers by adaptive dropout and generative in-filling

    Chun - Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classifiers by adaptive dropout and generative in-filling. CoRR, abs/1807.08024, 2018. URL http://arxiv.org/abs/1807.08024

    arXiv 2018

  6. [6]

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

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839--847, 2018. doi:10.1109/WACV.2018.00097

    arXiv 2018

  7. [7]

    Latent Tree Models for Hierarchical Topic Detection

    Peixian Chen, Nevin L. Zhang, Tengfei Liu, Leonard K. M. Poon, and Zhourong Chen. Latent tree models for hierarchical topic detection. CoRR, abs/1605.06650, 2016. URL http://arxiv.org/abs/1605.06650

    work page Pith review arXiv 2016

  8. [8]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248--255, 2009. doi:10.1109/CVPR.2009.5206848

    arXiv 2009

Show all 31 references
  1. [9]

    Ramaswamy

    Saurabh Desai and Harish G. Ramaswamy. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 972--980, 2020. doi:10.1109/WACV45572.2020.9093360

    2020

  2. [10]

    Understanding individual decisions of cnns via contrastive backpropagation

    Jindong Gu, Yinchong Yang, and Volker Tresp. Understanding individual decisions of cnns via contrastive backpropagation. CoRR, abs/1812.02100, 2018. URL http://arxiv.org/abs/1812.02100

    2018 arXiv

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385

    2015 arXiv

  4. [12]

    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. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Syste...

    2019

  5. [13]

    Weinberger

    Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. CoRR, abs/1608.06993, 2016. URL http://arxiv.org/abs/1608.06993

    2016 arXiv

  6. [14]

    Mohammad A. A. K. Jalwana, Naveed Akhtar, Mohammed Bennamoun, and Ajmal Mian. CAMERAS: enhanced resolution and sanity preserving class activation mapping for image saliency. CoRR, abs/2106.10649, 2021. URL https://arxiv.org/abs/2106.10649

    2021 arXiv

  7. [15]

    Layercam: Exploring hierarchical class activation maps for localization

    Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30: 0 5875--5888, 2021. doi:10.1109/TIP.2021.3089943

    2021

  8. [16]

    Measuring self-supervised representation quality for downstream classification using discriminative features, 2023

    Neha Kalibhat, Kanika Narang, Hamed Firooz, Maziar Sanjabi, and Soheil Feizi. Measuring self-supervised representation quality for downstream classification using discriminative features, 2023

    2023

  9. [17]

    Sunnie S. Y. Kim, Nicole Meister, Vikram V. Ramaswamy, Ruth Fong, and Olga Russakovsky. HIVE: evaluating the human interpretability of visual explanations. CoRR, abs/2112.03184, 2021. URL https://arxiv.org/abs/2112.03184

    2021 arXiv

  10. [18]

    The disagreement problem in explainable machine learning: A practitioner's perspective

    Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner's perspective. CoRR, abs/2202.01602, 2022. URL https://arxiv.org/abs/2202.01602

    2022 arXiv

  11. [19]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. URL https://api.semanticscholar.org/CorpusID:18268744

    2009

  12. [20]

    An evaluation of the human-interpretability of explanation

    Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Sam Gershman, and Finale Doshi - Velez. An evaluation of the human-interpretability of explanation. CoRR, abs/1902.00006, 2019. URL http://arxiv.org/abs/1902.00006

    1902 arXiv

  13. [21]

    Faithful and customizable explanations of black box models

    Himabindu Lakkaraju, Ece Kamar, Rich Caruana, and Jure Leskovec. Faithful and customizable explanations of black box models. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, AIES '19, page 131–138, New York, NY, USA, 2019. Association for Computing Ma...

    2019

  14. [22]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao - Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. CoRR, abs/2201.03545, 2022. URL https://arxiv.org/abs/2201.03545

    2022 arXiv

  15. [23]

    Sina Mohseni, Niloofar Zarei, and Eric D. Ragan. A multidisciplinary survey and framework for design and evaluation of explainable ai systems. ACM Trans. Interact. Intell. Syst., 11 0 (3–4), September 2021. ISSN 2160-6455. doi:10.1145/3387166. URL https://doi.org/10.1145/3387166

    2021 doi

  16. [24]

    RISE: randomized input sampling for explanation of black-box models

    Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: randomized input sampling for explanation of black-box models. CoRR, abs/1806.07421, 2018. URL http://arxiv.org/abs/1806.07421

    2018 arXiv

  17. [25]

    Evaluating the visualization of what a deep neural network has learned

    Wojciech Samek, Alexander Binder, Gr \' e goire Montavon, Sebastian Bach, and Klaus - Robert M \" u ller. Evaluating the visualization of what a deep neural network has learned. CoRR, abs/1509.06321, 2015. URL http://arxiv.org/abs/1509.06321

    2015 arXiv

  18. [26]

    Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Why did you say that? visual explanations from deep networks via gradient-based localization. CoRR, abs/1610.02391, 2016. URL http://arxiv.org/abs/1610.02391

    arXiv

  19. [27]

    A survey on explainable artificial intelligence (xai): Toward medical xai

    Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (xai): Toward medical xai. IEEE Transactions on Neural Networks and Learning Systems, 32 0 (11): 0 4793--4813, 2021. doi:10.1109/TNNLS.2020.3027314

    2021

  20. [28]

    Score-cam: Improved visual explanations via score-weighted class activation mapping

    Haofan Wang, Mengnan Du, Fan Yang, and Zijian Zhang. Score-cam: Improved visual explanations via score-weighted class activation mapping. CoRR, abs/1910.01279, 2019. URL http://arxiv.org/abs/1910.01279

    1910 arXiv

  21. [29]

    Weiyan Xie, Xiao-Hui Li, Zhi Lin, Leonard K. M. Poon, Caleb Chen Cao, and Nevin L. Zhang. Two-stage holistic and contrastive explanation of image classification, 2023

    2023

  22. [30]

    Inouye, and Pradeep Ravikumar

    Chih - Kuan Yeh, Cheng - Yu Hsieh, Arun Sai Suggala, David I. Inouye, and Pradeep Ravikumar. How sensitive are sensitivity-based explanations? CoRR, abs/1901.09392, 2019. URL http://arxiv.org/abs/1901.09392

    1901 arXiv

  23. [31]

    Learning deep features for discriminative localization

    Bolei Zhou, Aditya Khosla, \` A gata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. CoRR, abs/1512.04150, 2015. URL http://arxiv.org/abs/1512.04150

    2015 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.