Pith. sign in

REVIEW 4 major objections 3 minor 76 references

On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read This paper claims that noisy, unfaithful gradient explanations are two views of one spectral fact: the tail of a network's power spectrum. It proposes Expected Frequency and the explanation gap as a parameter-free way to measure and trade o

desk verdict Solid empirical story about smooth activations, but the spectral trade-off framework rests on an unproven equivalence that fails even in simple linear cases. read the letter →

arxiv 2508.10490 v1 pith:WFIQ3VFC submitted 2025-08-14 cs.LG cs.AIcs.CV

classification cs.LGcs.AIcs.CV
keywords gradient-basedexplanationsexplanationcomplexityfaithfulnessexpectedfrequencygappowerspectrumReLUnetworksspectralanalysis
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 tries to establish that the two standard complaints about gradient-based explanations — vanilla saliency maps look noisy, and smoothing methods fix the noise by betraying the model — are one phenomenon seen through a single lens: the tail of the network's power spectrum. The authors claim that on image-like data, the high-frequency tail of the network function is directly proportional to the high-frequency tail of its input gradient, so a summary statistic of that tail, the Expected Frequency, measures explanation complexity, and its change under a surrogate method, the 'explanation gap,' measures unfaithfulness. They further claim that smoothing the ReLU activation by convolving it with a Gaussian (approximated by SoftPlus) shortens the tail, producing simpler explanations and letting a designer choose where to sit on the complexity–faithfulness curve. If the claims hold, explainability stops being a matter of taste or heuristics: the trade-off becomes a measurable, adjustable spectral property of the architecture itself.

What carries the argument

The load-bearing objects are two spectral tails and the bridge between them: the Tail of the Power Spectrum (TPS) of the network $f$ and the Tail of the Spatial Power Spectrum (TSPS) of its input gradient $\nabla f$, whose proportional decay under high spatial autocorrelation is Theorem 1. On top of these sit two summary statistics: Expected Frequency $\mathrm{EF}(\mathbf{e}_f)=\int \omega S_{\mathbf{e}_f}(\omega)\,d\omega$ (Eq. 1), a one-number measure of explanation complexity, and the explanation gap, $G(f,\tilde f)\sim\Delta\mathrm{EF}:=|\mathrm{EF}(\nabla f)-\mathrm{EF}(\mathbf{e}_f)|$ (Eq. 7), a faithfulness proxy obtained by applying Parseval's theorem so the $L^2$ gap between origina

What would settle it

Train identical architectures on an image dataset with pixels randomly permuted within each image (destroying spatial autocorrelation) and compare VanillaGrad's expected frequency against the network's power-spectrum tail across the SoftPlus-to-ReLU parameter range. Theorem 1 predicts the two tails move together; if the proportionality breaks on such inputs, the autocorrelation assumption is doing the load-bearing work. A complementary check: on natural images, craft an explained sample whose pixel-value trajectory provably avoids all training trajectories and see whether EF still tracks the n

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a proportionality of tails: for a trained network $f$ on image-like data, the tail behavior of the power spectrum of $f$ is directly proportional to the tail of the spatial power spectrum of $\nabla f$ (Theorem 1). Because a network's Fourier transform is inaccessible, the gradient's spatial spectrum is read as the aperture onto the network's spectrum. From this the paper derives two quantities in one framework: Expected Frequency $\mathrm{EF}(\mathbf{e}_f)=\int \omega S_{\mathbf{e}_f}(\omega)\,d\omega$ for explanation complexity, and the explanation gap $G(f,\tilde f)\sim\Delta\mathrm{EF}=|\mathrm{EF}(\nabla f)-\mathrm{EF}(\mathbf{e}_f)|$ f

Load-bearing premise

That the input data and the features learned on it are highly autocorrelated in space (neighboring pixels vary slowly), and that the explained sample's pixel-value trajectory crosses a training sample's; the authors state in Appendix G that the autocorrelation property 'has not been theoretically proven yet,' so if either condition fails, the tail-proportionality theorem and the explanation-gap proxy built on it lose their foundation.

Editorial extensions

If this is right

  • Explanation complexity becomes measurable as a single number, EF, and faithfulness as its change, ΔEF, so post-hoc methods can be compared on both axes at once without extra hyperparameters.
  • Smoothing the activation function (SoftPlus parameterization of ReLU) demonstrably reduces EF; model designers can trade a little accuracy for substantially cleaner VanillaGrad explanations.
  • Surrogate-based methods act as low-pass filters: SmoothGrad and GradCAM reduce EF at a measurable faithfulness cost, with GradCAM showing the largest explanation gap across architectures.
  • The trade-off is continuous rather than a post-hoc versus ante-hoc dichotomy: one can achieve zero explanation gap at reduced complexity (VanillaGrad on a smooth network), making the two goals orthogonal axes rather than a binary choice.
  • The framework is shift- and rotation-invariant (Fourier basis plus power spectrum) and insensitive to gradient magnitude after rank normalization, addressing known confounds in pixel-removal faithfulness metrics.

Reading between the lines

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

  • My extension: if Theorem 1 survives closer scrutiny, the spectral tail becomes a candidate optimization objective for architecture search and activation design — networks could be selected or regularized directly for explanation complexity, which the paper only gestures at in its limitations.
  • My extension: the ΔEF proxy presumes surrogate methods act as low-pass filters; for methods that add or redistribute high-frequency content (e.g., Integrated Gradients' path dependence), ΔEF may understate the true explanation gap, so the faithfulness axis is only reliable within the low-pass family.
  • My extension: the framework predicts that any architectural choice known to flatten the learned function's spectrum — such as batch normalization, which the paper's ablation shows amplifies the tail — will shift EF in a predictable direction, giving a cheap testable fingerprint for new architectures, including transformers, where the paper notes its Laplace-kernel assumption breaks.
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

4 major / 3 minor

Summary. The paper proposes a spectral framework for analyzing gradient-based explanations. It defines Expected Frequency (EF), a summary statistic of the spatial power spectrum of input gradients, and argues that the tail of a network's power spectrum (TPS) is directly reflected in the tail of the spatial power spectrum of its gradient (TSPS). On this basis it introduces a measure of faithfulness, the 'explanation gap' G, and claims that G can be approximated by the absolute change in EF, ΔEF, induced by a surrogate explanation method. The authors further study a smooth parameterization of ReLU (SoftPlus, viewed as ReLU convolved with a Gaussian) and provide experiments across Fashion-MNIST, CIFAR-10, Imagenette, and ImageNet showing that smoother activations reduce EF and that post-hoc methods such as GradCAM and SmoothGrad have larger ΔEF. The central advertised contribution is a unified, parameter-free framework in which complexity and faithfulness can be measured and traded off.

Significance. If the framework were sound, it would provide a useful unifying spectral perspective on explanation complexity and faithfulness, and the empirical observation that smoother activations reduce gradient noise is plausible and well demonstrated across several datasets and ablations. The authors deserve credit for a broad experimental study (Figs. 2, 3, 7-9, 11; Tabs. 1-4), including ablations on depth, learning rate, input size, skip connections, and batch normalization, and for releasing code. However, the central theoretical bridge is not established. Specifically, the claimed equivalence G ~ ΔEF in Eq. (7) is not derived, is false in general, and is close to a definitional construct because ΔEF is defined as a difference of EF values. The main results therefore do not support the advertised trade-off framework.

major comments (4)
  1. [Sec. 4.4, Eq. (7) and Appendix E] Eq. (7) asserts G(f, f̃) ~ ΔEF(ef) := |EF(∇f) − EF(ef)|, but G and ΔEF are different functionals. By Parseval, G = ∫ ω²|F∇f − F ef|² dω, while ΔEF = |∫ω(|F∇f|² − |F ef|²)dω|. The former depends on the cross-spectrum (phase alignment) of ∇f and ef; the latter only on their marginal power spectra. Two explanations with identical power spectra but different phase (e.g., ef = −∇f, or a spatial mirror) give ΔEF = 0 while G > 0. Appendix E does not close this gap: it only shows that for a Laplace kernel both quantities scale as O(1/b) or O(b). Same leading-order asymptotics does not imply proportionality, and the appendix never computes G from the actual difference ∇f − ef. Thus the load-bearing bridge of the paper is unsupported.
  2. [Sec. 4.1 and Sec. 4.4] The claimed complexity-faithfulness trade-off is, to a large extent, definitional. Complexity is defined as EF in Eq. (1), and the faithfulness measure is then defined as ΔEF = |EF(∇f) − EF(ef)| in Eq. (7). Tables 1 and 2 therefore report the difference between two EF values as an 'explanation gap'; this is not an independent faithfulness metric. The paper does not validate ΔEF against any established faithfulness measure (e.g., pixel removal or sanity checks), so the trade-off is a relation between a quantity and its own variation rather than an empirical finding about faithfulness.
  3. [Theorem 1, Remark 6, Eq. (23)-(24), Appendix G] The proof of Theorem 1 rests on two assumptions that are acknowledged in the paper to be unproven or only heuristic: high spatial autocorrelation of inputs and learned features (Remark 6, Eq. (23)) and the existence of an intersection between the explained sample and a training sample in pixel-value space (Eq. (24)). Appendix G explicitly states that the autocorrelation property 'has not been theoretically proven yet.' Without these assumptions, Lemma 3 and the proportionality between the network TPS and the gradient TSPS collapse. The theorem is also stated informally and proved only in a drastically simplified setting (n=1, α=1, kernel expansion with the NTK replaced by a Laplace kernel), so it does not establish the claim for general trained networks.
  4. [Appendix F, Eq. (44)] The proof of Lemma 1 contains a mathematical error. The authors define gβ with precision β, i.e., variance 1/β. After convolving the activation with gβ, the covariance of the τ-transform inputs should be σ² + 1/β on the diagonal, not σ² + β as written in Eq. (44). The subsequent claim that increasing β makes the matrix 'closer to identity' is a consequence of this erroneous variance. With the correct variance, increasing β makes the distribution closer to the original ReLU case, not closer to identity. The empirical spectrum in Fig. 11 supports the intended conclusion, but the provided proof is not valid.
minor comments (3)
  1. [Throughout] There are several typos: 'whith' (Fig. 7 caption), 'hyperparamters' (Appendix B), 'an comprehensive' (Sec. 6), 'T est' (Fig. 5), 'hypothetize' (Sec. 3.2).
  2. [Sec. 4.3, Eq. (6)] The approximation that the low-frequency integral vanishes after surrogate smoothing is not justified. Low-pass surrogates do not generally leave low frequencies unchanged, and the threshold ω* is a free parameter. This step should be clarified or removed.
  3. [Sec. 5] The validation accuracy cap is described as a way to compare under similar training budgets, but it also means that the compared networks have different accuracy levels. The ablation in Fig. 5 addresses this, but the main text should state more explicitly that the EF reductions come with accuracy trade-offs.

Circularity Check

1 steps flagged · score 6.0 of 10

The faithfulness axis is defined as a difference of the complexity metric, making the advertised trade-off a mathematical identity rather than an empirical finding; the central G∼ΔEF proxy is asserted, not derived.

  1. self definitional [Sec. 4.4, Eq. (7); Tables 1 and 2]
    "More formally, we can use the absolute change in EF caused by an explanation method as a proxy for explanation gap, expressed as: G(f, ˜f ) ∼ ∆ EF(ef ) := | EF(∇f ) − EF(ef )| (7) ... This table shows the evaluation of expected frequency in Eq. (1) and explanation gap in Eq. (7)"

    EF is the paper's own complexity measure (Eq. 1). Eq. (7) defines the faithfulness proxy ΔEF as the absolute difference of two EF values, so for a fixed original model ΔEF is a deterministic function of the surrogate's EF. Consequently, any method that lowers EF automatically has a positive 'explanation gap'; the trade-off displayed in Fig. 1 and Tables 1–2 is a consequence of the definition, not an empirical discovery. The conclusion's claim that EF and ΔEF are 'two orthogonal axes' is contradicted by construction, since ΔEF is derived from the same expected-frequency statistic. The true gap G defined in Eq. (2) involves the L2 distance between gradients, which depends on cross-spectra/phase, and is not equal to a difference of marginal power-spectrum averages; the paper does not derive G

full rationale

The paper contains substantial independent content in Theorem 1 and Lemma 1: the tail of the spatial power spectrum of the gradient is related to the tail of the network's power spectrum under explicit assumptions (Eqs. 23–24), and smoothing ReLU by Gaussian convolution is shown to affect kernel spectral decay. These results are not circular. The main circularity is in the complexity–faithfulness framework itself. The faithfulness measure used for all empirical claims is ΔEF, which is defined as a difference of EF values. Since EF is the complexity measure, any surrogate that reduces EF necessarily increases ΔEF; the reported 'trade-off' is therefore a definitional identity rather than evidence about faithfulness. The paper also states in Appendix G that the key autocorrelation assumption 'has not been theoretically proven yet,' which is an unsupported premise, and Appendix E's derivation of Eq. (7) only analyzes kernel power spectra, never computing G from the actual gradient difference ∇f−∇f̃. These issues make the central trade-off claim only partially supported, but they are not self-citation circularity. The self-citation to [41] for the low-pass-filter view of explanation methods is a normal external reference and is not load-bearing in a circular way.

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

The framework rests on a kernel view of trained networks, high spatial autocorrelation of images, and an unproven replacement of NTK by the Laplace kernel. The metrics themselves involve hand-chosen parameters (β, accuracy cap, frequency split) and normalization choices. No new physical or ontological entities are postulated; EF, ΔEF and the smooth parameterization are definitions and design choices rather than entities.

free parameters (3)
  • Gaussian precision β (SoftPlus smoothness) = varied over 0.1 to 50 depending on dataset
    Hand-chosen hyperparameter interpolating between SoftPlus-smoothed ReLU and standard ReLU; all frequency-complexity results are reported as a function of β.
  • Validation accuracy cap = 60% Imagenette, 70% CIFAR-10, 80% Fashion-MNIST
    Chosen by hand to equalize training budgets and prevent initialization bias; the ablation without the cap shows the spectral trend is less clear (Fig. 6).
  • Frequency threshold ω* = not specified
    Introduced in Eq. (6) to split the explanation gap into low and high frequencies; the low-frequency integral is then assumed to vanish, but no value or sensitivity analysis is given.
assumptions (4)
  • domain assumption Input and learned features have high spatial autocorrelation (concentration of spatial power spectrum near zero).
    Explicitly assumed in Remark 6 (Eq. 23) and Appendix G; authors state this is intuitive for images but not theoretically proven. Without it, Lemma 3 and Theorem 1 lose the connection between network TPS and gradient TSPS.
  • domain assumption There exists at least one intersection between the explained sample and a training sample in pixel-value space (Eq. 24).
    Needed for Lemma 3 to linearize Δ(τ) and obtain the O(ω²|k̂|²) tail; the authors argue absence is improbable but provide no proof.
  • ad hoc to paper The trained network can be represented by a kernel expansion with its NTK, and the NTK can be replaced by the Laplace kernel without loss of generality for tail analysis.
    This substitution is asserted in Appendix C and used in Lemma 1 and the Eq. (7) derivation; only qualitative similarity between NTK and Laplace kernel is cited.
  • standard math Continuity and absolute integrability of the network function so that Fourier representations exist.
    Standard Fourier analysis assumption stated in Sec. 3.1; appropriate for continuous piecewise-linear ReLU networks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations." pith.science (2026). https://pith.science/paper/WFIQ3VFC

@misc{pith2026250810490,
  author       = {Pith},
  title        = {Pith review of: On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFIQ3VFC}},
  note         = {Machine review of arXiv:2508.10490}
}
read the original abstract

ReLU networks, while prevalent for visual data, have sharp transitions, sometimes relying on individual pixels for predictions, making vanilla gradient-based explanations noisy and difficult to interpret. Existing methods, such as GradCAM, smooth these explanations by producing surrogate models at the cost of faithfulness. We introduce a unifying spectral framework to systematically analyze and quantify smoothness, faithfulness, and their trade-off in explanations. Using this framework, we quantify and regularize the contribution of ReLU networks to high-frequency information, providing a principled approach to identifying this trade-off. Our analysis characterizes how surrogate-based smoothing distorts explanations, leading to an ``explanation gap'' that we formally define and measure for different post-hoc methods. Finally, we validate our theoretical findings across different design choices, datasets, and ablations.

Figures

Figures reproduced from arXiv: 2508.10490 by the authors.

Figure 1
Figure 1. Exploring the Faithfulness-Complexity Trade-Off. This figure shows the trade-off typically made between faithfulness and complexity in post-hoc explanation methods. (b) presents VanillaGrad (VG) explanations for a ReLU network, which appear grainy due to the network’s reliance on high-frequency information. To mitigate this, post-hoc methods like SmoothGrad (SG), shown in (c), remove high-frequency components by cre… view at source ↗
Figure 2
Figure 2. ReLU Affects Gradient TSPS via Network’s TPS. This figure presents a practical implication of our theory on the tail behavior of the spatial power spectrum in relation to function sharpness, as formalized in Lemma 1. The plots depict the TSPS of input-gradients of networks trained with identical architectures and hyperparameters, varying only in smoothness parameter of ReLU (indicated in the legend) and datasets: (a… view at source ↗
Figure 3
Figure 3. Impact of Smoothness Parameter on Expected Frequency. This figure shows EF, defined in Eq. (1), used as a summary statistic of the frequency content of explanations in the spatial domain with respect to the smoothness parameter β in a (log) x-axis. To obtain the 95% CI, we repeated experiments using 10 seeds for each dataset: (a) Fashion MNIST, (b) CIFAR-10, (c) Imagenette (122×122), and (d) Imagenette (224 × 224). … view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Overview of Key Theoretical Connections. A graphical overview of our contributions, which is based on prior works on spectral analysis of gradient-based explanation methods. The diagram illustrates the conceptual flow of the narrative presented in this paper. From TPS …
Figure 5
Figure 5. Figure 5: Ablation Study: Impact of Smoothness Parameter on Validation Accuracy. This figure presents an ablation study of our decision to impose an accuracy cap as an early stopping mech￾anism on Imagenette (224 × 224). By relaxing this constraint, we train smooth parameterizat…
Figure 6
Figure 6. Figure 6: Ablation Study: Impact of Validation Accuracy Cap and Learning Rate on the Spatial Power Spectrum. This figure presents an ablation study on our decision to impose a validation accuracy cap as an early stopping mechanism on Imagenette (224 × 224) across different learn…
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Ablation Study: Impact of Input Size on the Spatial Power Spectrum. This figure presents an ablation study, isolating the effect of input size by training multiple models on different versions of the Imagenette dataset with varying input resolutions. The input sizes ra…
Figure 9
Figure 9. Figure 9: Ablation Study: Effect of Skip Connections and Batch Normalization on the Spatial Power Spectrum Tail. This figure illustrates the impact of skip connections and batch normal￾ization on the tail of the spatial power spectrum on Imagenette (224 × 224). In (a) the power …
Figure 10
Figure 10. Figure 10: Illustration of Kernel Sharpness on the Spatial Power Spectrum Tail. A cartoon illustration depicts the theory regarding the effect of sharpness of the learned features on the tail of the power spectrum. (a) and (b): The x-axis represents a simplified version of spati…
Figure 11
Figure 11. Figure 11: Spectral Decay of the Empirical NTK Across Smooth Parameterizations of ReLU. This figure depicts the spectral decay of the empirical neural tangent kernel NTK, plotting the eigenvalues (x-axis) against their normalized magnitudes (y￾axis). This empirical evaluation su…
Figure 12
Figure 12. Figure 12: Impact of Low Spatial Autocorrelation on Spatial Power Spectrum of Gradient. This figure illustrates the impact of low spatial autocorrelation in the input and, consequently, in the learned features on the tail behavior of the spatial spectral density of the gradient.…
Figure 13
Figure 13. Figure 13: Sample Visualization of Explanations Outputs After Inverse Transformation Normalization. This figure presents a visualization of the outputs from various explanation methods after applying the inverse transformation method. This normalization technique aligns differen…
Figure 14
Figure 14. Figure 14: Sample Visualization of VanillaGrad Explanations Across Smooth ReLU Parameterizations. This figure presents addi￾tional samples from trained networks with different smooth parameterizations of ReLU, denoted by SP(β). The goal is to illustrate how the reliance on high-…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 42 canonical work pages

  1. [1]

    Sanity Checks for Saliency Maps, 2020

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Good- fellow, Moritz Hardt, and Been Kim. Sanity Checks for Saliency Maps, 2020. arXiv:1810.03292 [cs, stat]. 1, 12

  2. [2]

    Genevera I. Allen. Automatic Feature Selection via Weighted Kernels and Regularization. Journal of Compu- tational and Graphical Statistics, 22(2):284–299, 2013. 14

  3. [3]

    On Pixel-Wise Explanations for Non-Linear Classi- fier Decisions by Layer-Wise Relevance Propagation

    Sebastian Bach, Alexander Binder, Gr ´egoire Montavon, Frederick Klauschen, Klaus-Robert M ¨uller, and Wojciech Samek. On Pixel-Wise Explanations for Non-Linear Classi- fier Decisions by Layer-Wise Relevance Propagation. PLOS ONE, 10(7):e0130140, 2015. 2

  4. [4]

    David Balduzzi, Marcus Frean, Lennox Leary, J. P. Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The Shattered Gradients Problem: If resnets are the answer, then what is the question?, 2018. arXiv:1702.08591 [cs, stat]. 12

  5. [5]

    Feature learning as alignment: a structural prop- erty of gradient descent in non-linear neural networks, 2024

    Daniel Beaglehole, Ioannis Mitliagkas, and Atish Agar- wala. Feature learning as alignment: a structural prop- erty of gradient descent in non-linear neural networks, 2024. arXiv:2402.05271 [cs, stat]. 14, 16

  6. [6]

    Umang Bhatt, Adrian Weller, and Jos ´e M. F. Moura. Eval- uating and Aggregating Feature-based Model Explanations,

  7. [7]

    Deep Equals Shal- low for ReLU Networks in Kernel Regimes, 2021

    Alberto Bietti and Francis Bach. Deep Equals Shal- low for ReLU Networks in Kernel Regimes, 2021. arXiv:2009.14397 [stat]. 8, 12, 15

  8. [8]

    On the Inductive Bias of Neural Tangent Kernels, 2019

    Alberto Bietti and Julien Mairal. On the Inductive Bias of Neural Tangent Kernels, 2019. arXiv:1905.12173. 14, 16

Show all 76 references
  1. [9]

    Layer-wise Rel- evance Propagation for Neural Networks with Local Renor- malization Layers, 2016

    Alexander Binder, Gr ´egoire Montavon, Sebastian Bach, Klaus-Robert M¨uller, and Wojciech Samek. Layer-wise Rel- evance Propagation for Neural Networks with Local Renor- malization Layers, 2016. arXiv:1604.00825 [cs]. 7, 14

  2. [10]

    Kirill Bykov, Anna Hedstr ¨om, Shinichi Nakajima, and Marina M.-C. H ¨ohne. NoiseGrad: Enhancing Explana- tions by Introducing Stochasticity to Model Weights, 2022. arXiv:2106.10185 [cs]. 2, 3

  3. [11]

    B-cos Net- works: Alignment is All We Need for Interpretability, 2022

    Moritz B ¨ohle, Mario Fritz, and Bernt Schiele. B-cos Net- works: Alignment is All We Need for Interpretability, 2022. 2

  4. [12]

    Towards the Spectral bias Alleviation by Normaliza- tions in Coordinate Networks, 2024

    Zhicheng Cai, Hao Zhu, Qiu Shen, Xinran Wang, and Xun Cao. Towards the Spectral bias Alleviation by Normaliza- tions in Coordinate Networks, 2024. arXiv:2407.17834 [cs]. 12

  5. [13]

    Can I Trust the Explainer? Verifying Post-hoc Explanatory Methods, 2019

    Oana-Maria Camburu, Eleonora Giunchiglia, Jakob Foer- ster, Thomas Lukasiewicz, and Phil Blunsom. Can I Trust the Explainer? Verifying Post-hoc Explanatory Methods, 2019. arXiv:1910.02065 [cs]. 2

  6. [14]

    Janizek, Scott Lundberg, and Su- In Lee

    Hugh Chen, Joseph D. Janizek, Scott Lundberg, and Su- In Lee. True to the Model or True to the Data?, 2020. arXiv:2006.16234 [cs, stat]. 1

  7. [15]

    Wainwright, and Michael I

    Jianbo Chen, Mitchell Stern, Martin J. Wainwright, and Michael I. Jordan. Kernel Feature Selection via Conditional Covariance Minimization, 2018. arXiv:1707.01164 [stat]. 14

  8. [16]

    Explaining by Removing: A Unified Framework for Model Explanation,

    Ian Covert, Scott Lundberg, and Su-In Lee. Explaining by Removing: A Unified Framework for Model Explanation,

  9. [17]

    Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity, 2017

    Amit Daniely, Roy Frostig, and Yoram Singer. Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity, 2017. arXiv:1602.05897. 14

  10. [18]

    ImageNet: A large-scale hierarchical im- age database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical im- age database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-

  11. [19]

    On the lipschitz constant of deep networks and double de- scent

    Matteo Gamba, Hossein Azizpour, and Marten Bjorkman. On the lipschitz constant of deep networks and double de- scent. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023. BMV A,

  12. [20]

    On the Similarity between the Laplace and Neural Tangent Kernels, 2020

    Amnon Geifman, Abhay Yadav, Yoni Kasten, Meirav Galun, David Jacobs, and Ronen Basri. On the Similarity between the Laplace and Neural Tangent Kernels, 2020. 14, 16, 20

  13. [21]

    Large-scale Nonlinear Variable Selection via Kernel Random Features, 2018

    Magda Gregorov ´a, Jason Ramapuram, Alexandros Kalousis, and St ´ephane Marchand-Maillet. Large-scale Nonlinear Variable Selection via Kernel Random Features, 2018. arXiv:1804.07169 [cs]. 14

  14. [22]

    Spectral analysis based on signal depen- dent transformation

    Modris Greitans. Spectral analysis based on signal depen- dent transformation. Proc. SMMSP 2005 , pages 179–184,

  15. [23]

    Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations, 2022

    Tessa Han, Suraj Srinivas, and Himabindu Lakkaraju. Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations, 2022. arXiv:2206.01254 [cs]. 1

  16. [24]

    Gradient Noise Convolution (GNC): Smooth- ing Loss Function for Distributed Large-Batch SGD, 2019

    Kosuke Haruki, Taiji Suzuki, Yohei Hamakawa, Takeshi Toda, Ryuji Sakai, Masahiro Ozawa, and Mitsuhiro Kimura. Gradient Noise Convolution (GNC): Smooth- ing Loss Function for Distributed Large-Batch SGD, 2019. arXiv:1906.10822 [cs, stat]. 12

  17. [25]

    Deep Residual Learning for Image Recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition, 2015. arXiv:1512.03385 [cs]. 7

  18. [26]

    A Benchmark for Interpretability Methods in Deep Neural Networks, 2019

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A Benchmark for Interpretability Methods in Deep Neural Networks, 2019. arXiv:1806.10758 [cs, stat]. 2

  19. [27]

    Zejiang Hou and S. Y . Kung. A Kernel Discriminant In- formation Approach to Non-linear Feature Selection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–10, 2019. ISSN: 2161-4407. 14

  20. [28]

    Neural Tangent Kernel: Convergence and Generalization in Neural Networks, 2020

    Arthur Jacot, Franck Gabriel, and Cl ´ement Hongler. Neural Tangent Kernel: Convergence and Generalization in Neural Networks, 2020. arXiv:1806.07572. 1, 5, 13

  21. [29]

    Simoncelli, and St´ephane Mallat

    Zahra Kadkhodaie, Florentin Guth, Eero P. Simoncelli, and St´ephane Mallat. Generalization in diffusion models arises from geometry-adaptive harmonic representations, 2024. arXiv:2310.02557 [cs]. 2, 23

  22. [30]

    Sriperumbudur

    Motonobu Kanagawa, Philipp Hennig, Dino Sejdinovic, and Bharath K. Sriperumbudur. Gaussian Processes and Ker- nel Methods: A Review on Connections and Equivalences,

  23. [31]

    Guided In- tegrated Gradients: An Adaptive Path Method for Removing Noise, 2021

    Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided In- tegrated Gradients: An Adaptive Path Method for Removing Noise, 2021. arXiv:2106.09788 [cs]. 2

  24. [32]

    Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps,

    Beomsu Kim, Junghoon Seo, SeungHyun Jeon, Jamyoung Koo, Jeongyeol Choe, and Taegyun Jeon. Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps,

  25. [33]

    Sch ¨utt, Sven D ¨ahne, Dumitru Er- han, and Been Kim

    Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Max- imilian Alber, Kristof T. Sch ¨utt, Sven D ¨ahne, Dumitru Er- han, and Been Kim. The (Un)reliability of saliency methods,

  26. [34]

    Captum: A unified and generic model interpretability library for PyTorch, 2020

    Narine Kokhlikyan, Vivek Miglani, Miguel Martin, Edward Wang, Bilal Alsallakh, Jonathan Reynolds, Alexander Mel- nikov, Natalia Kliushkina, Carlos Araya, Siqi Yan, and Orion Reblitz-Richardson. Captum: A unified and generic model interpretability library for PyTorch, 2020. 8

  27. [35]

    Cartoon Explanations of Image Classifiers,

    Stefan Kolek, Duc Anh Nguyen, Ron Levie, Joan Bruna, and Gitta Kutyniok. Cartoon Explanations of Image Classifiers,

  28. [36]

    Explaining Image Clas- sifiers with Multiscale Directional Image Representation,

    Stefan Kolek, Robert Windesheim, Hector Andrade Loarca, Gitta Kutyniok, and Ron Levie. Explaining Image Clas- sifiers with Multiscale Directional Image Representation,

  29. [37]

    Partial order: Find- ing Consensus among Uncertain Feature Attributions, 2023

    Gabriel Laberge, Yann Pequignot, Foutse Khomh, Mario Marchand, and Alexandre Mathieu. Partial order: Find- ing Consensus among Uncertain Feature Attributions, 2023. arXiv:2110.13369 [cs]. 17

  30. [38]

    Gershman, and Finale Doshi-Velez

    Isaac Lage, Emily Chen, Jeffrey He, Menaka Narayanan, Been Kim, Samuel J. Gershman, and Finale Doshi-Velez. Human Evaluation of Models Built for Interpretability. Pro- ceedings of the AAAI Conference on Human Computation and Crowdsourcing, 7:59–67, 2019. 2

  31. [39]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural informa- tion processing systems, 30, 2017. 17

  32. [40]

    Harmonics of Learning: Univer- sal Fourier Features Emerge in Invariant Networks, 2024

    Giovanni Luca Marchetti, Christopher Hillar, Danica Kragic, and Sophia Sanborn. Harmonics of Learning: Univer- sal Fourier Features Emerge in Invariant Networks, 2024. arXiv:2312.08550 [cs]. 2, 23

  33. [41]

    arXiv:2110.03485 [cs]. 2

  34. [42]

    Characterizing the Spectrum of the NTK via a Power Series Expansion, 2023

    Michael Murray, Hui Jin, Benjamin Bowman, and Guido Montufar. Characterizing the Spectrum of the NTK via a Power Series Expansion, 2023. arXiv:2211.07844 [cs]. 21

  35. [43]

    Abolafia, Jeffrey Pennington, and Jascha Sohl-Dickstein

    Roman Novak, Yasaman Bahri, Daniel A. Abolafia, Jeffrey Pennington, and Jascha Sohl-Dickstein. Sensitivity and Gen- eralization in Neural Networks: an Empirical Study, 2018. arXiv:1802.08760 [cs, stat]. 19

  36. [44]

    Mathematical theory of deep learning, 2024

    Philipp Petersen and Jakob Zech. Mathematical theory of deep learning, 2024. arXiv:2407.18384 [cs]. 8

  37. [45]

    Hamprecht, Yoshua Bengio, and Aaron Courville

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron Courville. On the Spectral Bias of Neural Networks,

  38. [46]

    How Reli- able and Stable are Explanations of XAI Methods?, 2024

    Jos ´e Ribeiro, Lucas Cardoso, Vitor Santos, Eduardo Car- valho, N ´ıkolas Carneiro, and Ronnie Alves. How Reli- able and Stable are Explanations of XAI Methods?, 2024. arXiv:2407.03108 [cs]. 1

  39. [47]

    A case for new neural network smooth- ness constraints, 2021

    Mihaela Rosca, Theophane Weber, Arthur Gretton, and Shakir Mohamed. A case for new neural network smooth- ness constraints, 2021. arXiv:2012.07969 [cs, stat]. 1, 5

  40. [48]

    On Spectral Properties of Gradient-Based Explanation Meth- ods

    Amir Mehrpanah, Erik Englesson, and Hossein Azizpour. On Spectral Properties of Gradient-Based Explanation Meth- ods. In Computer Vision – ECCV 2024 , pages 282–299, Cham, 2025. Springer Nature Switzerland. 1, 6, 7, 12, 13

  41. [49]

    Best of both worlds: local and global explanations with human-understandable concepts,

    Jessica Schrouff, Sebastien Baur, Shaobo Hou, Diana Mincu, Eric Loreaux, Ralph Blanes, James Wexler, Alan Karthike- salingam, and Been Kim. Best of both worlds: local and global explanations with human-understandable concepts,

  42. [50]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. International Journal of Computer Vision, 128(2):336–359, 2020. arXiv:1610.02391 [c...

  43. [51]

    Not Just a Black Box: Learning Important Features Through Propagating Activation Differences, 2017

    Avanti Shrikumar, Peyton Greenside, Anna Shcherbina, and Anshul Kundaje. Not Just a Black Box: Learning Important Features Through Propagating Activation Differences, 2017. arXiv:1605.01713 [cs]. 2, 7, 14

  44. [52]

    Simon, Sajant Anand, and Michael R

    James B. Simon, Sajant Anand, and Michael R. DeWeese. Reverse Engineering the Neural Tangent Kernel, 2022. arXiv:2106.03186 [cs]. 14, 20

  45. [53]

    arXiv:1806.08734 [cs, stat]. 2, 3

  46. [54]

    SmoothGrad: removing noise by adding noise, 2017

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Vi ´egas, and Martin Wattenberg. SmoothGrad: removing noise by adding noise, 2017. arXiv:1706.03825 [cs, stat]. 1, 2, 3, 5, 7, 14

  47. [55]

    Striving for Simplicity: The All Convolutional Net, 2015

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity: The All Convolutional Net, 2015. arXiv:1412.6806 [cs]. 2, 7, 14

  48. [56]

    Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead

    Cynthia Rudin. Stop explaining black box machine learn- ing models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206–215,

  49. [57]

    One pixel attack for fooling deep neural networks

    Jiawei Su, Danilo Vasconcellos Vargas, and Sakurai Kouichi. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828–841,

  50. [58]

    arXiv:2106.08641 [cs]. 2

  51. [59]

    On the Structural Sensitivity of Deep Convolutional Networks to the Directions of Fourier Basis Functions, 2019

    Yusuke Tsuzuku and Issei Sato. On the Structural Sensitivity of Deep Convolutional Networks to the Directions of Fourier Basis Functions, 2019. arXiv:1809.04098 [cs]. 2, 3

  52. [60]

    From Flexibility to Manipulation: The Slip- pery Slope of XAI Evaluation, 2024

    Kristoffer Wickstrøm, Marina Marie-Claire H ¨ohne, and Anna Hedstr¨om. From Flexibility to Manipulation: The Slip- pery Slope of XAI Evaluation, 2024. arXiv:2412.05592 [cs]. 1

  53. [61]

    Feature Importance Ranking for Deep Learning, 2020

    Maksymilian Wojtas and Ke Chen. Feature Importance Ranking for Deep Learning, 2020. arXiv:2010.08973 [cs]. 17

  54. [62]

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps, 2014

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisser- man. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps, 2014. arXiv:1312.6034 [cs]. 2, 7, 14

  55. [63]

    Visual Transform- ers: Token-based Image Representation and Processing for Computer Vision, 2020

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual Transform- ers: Token-based Image Representation and Processing for Computer Vision, 2020. arXiv:2006.03677 [cs]. 7, 14

  56. [64]

    Benchmarking Attri- bution Methods with Relative Feature Importance, 2019

    Mengjiao Yang and Been Kim. Benchmarking Attri- bution Methods with Relative Feature Importance, 2019. arXiv:1907.09701 [cs, stat]. 2 A. Explanation Methods are Low-pass Filters This section builds on key insights from studies focused on the spectral analysis of gradient-based...

  57. [65]

    Petre Stoica and Randolph L. Moses. Spectral analysis of signals. Pearson Prentice Hall Upper Saddle River, NJ, 2005. 8, 17

  58. [67]

    arXiv:1710.08864 [cs]. 1, 5

  59. [68]

    Axiomatic Attribution for Deep Networks, 2017

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic Attribution for Deep Networks, 2017. arXiv:1703.01365 [cs]. 2, 7, 14

  60. [72]

    Wright and Joseph E

    Matthew A. Wright and Joseph E. Gonzalez. Transform- ers are Deep Infinite-Dimensional Non-Mercer Binary Ker- nel Machines, 2021. arXiv:2106.01506 [cs]. 14

  61. [75]

    In this degenerate case, no root is present

    No Intersection: If there is no τ for which ∆(τ ) = 0, then a linear approximation yields ∆(τ ) = α with α ̸= 0 . In this degenerate case, no root is present

  62. [76]

    In a neighborhood of such a point, by a first-order Taylor expansion, ∆(τ ) ≈ α(τ − τ ∗), for some α ̸= 0 and τ close to τ ∗

    Intersection(s) Exist: There exists at least one τ ∗ satis- fying ∆(τ ∗) = 0. In a neighborhood of such a point, by a first-order Taylor expansion, ∆(τ ) ≈ α(τ − τ ∗), for some α ̸= 0 and τ close to τ ∗. This approximation ensures that the linearized ∆ has a root at τ = τ ∗, c...

  63. [2017]

    arXiv:1711.00867 [cs, stat]. 1

  64. [2018]

    arXiv:1807.02582 [cs, stat]. 14

  65. [2019]

    arXiv:1902.04893 [cs, stat]. 2

  66. [2020]

    arXiv:2005.00631 [cs]. 1, 2, 3

  67. [2022]

    arXiv:2011.14878 [cs, stat]. 6

  68. [2023]

    arXiv:2211.12857 [cs]. 2

Pith tools

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