pith. sign in

arxiv: 2605.16905 · v1 · pith:NQZJRTWSnew · submitted 2026-05-16 · 💻 cs.LG · cs.CV

AIM: Adversarial Information Masking for Faithfulness Evaluation of Saliency Maps

Chia-Ying Hsieh , Hsin-Yuan Fang , Chun-Shu Wei This is my paper

Pith reviewed 2026-05-19 20:20 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords saliency mapsfaithfulness evaluationadversarial maskingneural network interpretabilitypost-hoc explanationsmasking operatorsfeature attribution
0
0 comments X

The pith

AIM uses adversarial feature replacement to evaluate saliency map faithfulness with less masking bias.

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

The paper proposes Adversarial Information Masking as a way to test whether saliency maps correctly identify the features that drive a neural network's predictions. Standard evaluations mask features in saliency order and watch for drops in model performance, yet the choice of masking operator itself can create artifacts or leave residual signals that distort the results. AIM instead generates an adversarial version of each input and substitutes the selected features with values drawn from that version, then tracks degradation under both forward and reverse masking orders. Experiments across image, audio, and EEG data indicate that this approach lowers bias relative to zero or interpolation masking and exposes differences between signed and unsigned attributions that depend on the data type. A sympathetic reader would care because more reliable faithfulness checks would let practitioners select explanation methods they can actually trust when interpreting models used in high-stakes decisions.

Core claim

AIM replaces selected features with values from an adversarial counterpart of the input and compares degradation under complementary masking orders, yielding lower random-attribution bias and more stable faithfulness rankings than zero or interpolation masking while revealing modality-dependent differences between signed and unsigned attributions.

What carries the argument

Adversarial Information Masking (AIM), a saliency-guided framework that performs feature replacement using an adversarial counterpart of the input to isolate the effect of the saliency ordering from masking artifacts.

If this is right

  • Random-attribution bias drops compared with zero and interpolation masking across the tested modalities.
  • Rankings of explanation methods by faithfulness become more stable under the new evaluation.
  • Signed and unsigned attributions show different reliability patterns that vary by image, audio, or EEG data.
  • The method supplies a concrete way to check both saliency-map quality and masking-operator reliability in one procedure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Teams working on safety-critical models could adopt AIM to decide which saliency technique to deploy for a given data type.
  • The same adversarial-replacement idea might be tested on other post-hoc interpretation tools such as concept activation vectors.
  • Generating the adversarial counterpart without task-specific tuning would make the approach easier to apply across new domains.

Load-bearing premise

An adversarial counterpart of the input can be generated so that replacing the chosen features removes predictive information without introducing new confounding artifacts or residual signals.

What would settle it

A controlled test in which AIM-masked inputs still produce large performance changes that cannot be attributed to the saliency order, or in which the adversarial replacement visibly reintroduces predictive information detectable by the model.

Figures

Figures reproduced from arXiv: 2605.16905 by Chia-Ying Hsieh, Chun-Shu Wei, Hsin-Yuan Fang.

Figure 1
Figure 1. Figure 1: Overview of the proposed AIM framework. A post-hoc explanation method first produces a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Image-domain examples. AIM selects SGC++, while interpolation-based masking and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Audio spectrogram examples. AIM selects SGA, while interpolation-based masking selects GD and zero masking se￾lects SG. Aggregated mel-bin profiles show frequency-selective saliency patterns [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: LeRF curves for all explanation methods on MSOS using CNN14. Left: AIM; middle: [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustrative effect of absolute-value transformation on a sinusoidal saliency signal. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Post-hoc saliency methods are widely used to interpret deep neural networks, but their faithfulness is difficult to evaluate reliably. Existing evaluations mask features according to saliency-induced feature ordering and measure performance degradation, but this degradation can be confounded by the masking operator: zero masking may create out-of-distribution artifacts, while interpolation-based masking may preserve residual predictive information. We propose Adversarial Information Masking (AIM), a saliency-guided adversarial feature replacement framework for evaluating both saliency-map faithfulness and masking-operator reliability. AIM replaces selected features with values from an adversarial counterpart of the input and compares degradation under complementary masking orders. We assess reliability using random-attribution bias and stability of explanation-method faithfulness rankings. Experiments on image, audio, and EEG tasks suggest that AIM reduces masking-induced bias compared with zero and interpolation-based masking, while revealing modality-dependent differences between signed and unsigned attributions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes Adversarial Information Masking (AIM), a saliency-guided framework that replaces selected input features with values drawn from an adversarially generated counterpart to evaluate post-hoc saliency map faithfulness. It contrasts performance degradation under complementary masking orders against zero-masking and interpolation baselines, measuring reliability via random-attribution bias and stability of explanation-method rankings. Experiments on image, audio, and EEG tasks are reported to show reduced masking-induced bias and modality-dependent differences between signed and unsigned attributions.

Significance. If the central claim holds, AIM would address a recognized weakness in faithfulness evaluation by providing a masking operator that more cleanly isolates the removal of predictive information. The multi-modal scope is a positive feature that could reveal domain-specific behaviors of saliency methods. The work would be of moderate significance to the interpretability community provided the adversarial replacement step is shown to avoid residual signals or new artifacts.

major comments (2)
  1. [Method] The generation procedure for the adversarial counterpart (optimization target, constraints, and stopping criteria) is described at a high level only. This is load-bearing for the central claim because any leakage of predictive information or introduction of confounding artifacts would render the reported reduction in random-attribution bias and the stability comparisons inconclusive relative to the zero and interpolation baselines.
  2. [Experiments] No quantitative results, error bars, dataset sizes, or statistical tests are supplied in the abstract or summary of experiments. Without these details it is impossible to judge the magnitude or reliability of the claimed bias reduction and ranking stability across the three modalities.
minor comments (2)
  1. [Method] Clarify the precise definition of 'complementary masking orders' and how they are constructed for each modality.
  2. The distinction between signed and unsigned attributions should be formalized with explicit notation or equations to avoid ambiguity when discussing modality-dependent differences.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review. We respond to each major comment in turn and describe the changes we will make to the manuscript.

read point-by-point responses

  1. Referee: [Method] The generation procedure for the adversarial counterpart (optimization target, constraints, and stopping criteria) is described at a high level only. This is load-bearing for the central claim because any leakage of predictive information or introduction of confounding artifacts would render the reported reduction in random-attribution bias and the stability comparisons inconclusive relative to the zero and interpolation baselines.

    Authors: We concur that the adversarial counterpart generation is described at a high level in the current version. Since this is central to validating the reduction in bias, we will revise the method description to include full details on the optimization target, constraints, and stopping criteria. This will allow readers to assess potential issues with leakage or artifacts more thoroughly. revision: yes

  2. Referee: [Experiments] No quantitative results, error bars, dataset sizes, or statistical tests are supplied in the abstract or summary of experiments. Without these details it is impossible to judge the magnitude or reliability of the claimed bias reduction and ranking stability across the three modalities.

    Authors: The referee correctly notes the absence of quantitative details in the abstract and experiment summary. We will update these sections in the revised manuscript to include representative quantitative results, error bars, dataset sizes, and information on statistical tests supporting the bias reduction and stability claims. revision: yes

Circularity Check

0 steps flagged

No circularity in AIM derivation; new masking operator and evaluation are independent of inputs

full rationale

The paper introduces Adversarial Information Masking (AIM) to address confounding in existing saliency faithfulness evaluations by replacing masked features with values from an adversarially generated counterpart input. This is presented as a novel framework compared against zero and interpolation baselines, with reliability assessed via random-attribution bias and ranking stability across image, audio, and EEG experiments. No derivation step reduces by construction to a fitted parameter, self-definition, or load-bearing self-citation; the central proposal and empirical claims rest on the explicit design of the replacement operator and direct task evaluations rather than presupposing outcomes from prior fitted results or renamed patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on the unstated premise that adversarial inputs can be constructed to isolate saliency effects cleanly; no explicit free parameters, axioms, or invented entities are detailed.

pith-pipeline@v0.9.0 · 5682 in / 1085 out tokens · 38331 ms · 2026-05-19T20:20:05.738400+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

76 extracted references · 76 canonical work pages · 5 internal anchors

  1. [1]

    Visualizing and understanding convolutional networks

    Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision (ECCV), pages 818–833. Springer, 2014

    work page 2014

  2. [2]

    Evaluating the visualization of what a deep neural network has learned.IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660–2673, 2016

    Wojciech Samek, Alexander Binder, Gregoire Montavon, Sebastian Lapuschkin, and Klaus- Robert Muller. Evaluating the visualization of what a deep neural network has learned.IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660–2673, 2016

    work page 2016

  3. [3]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, volume 30, 2017

    work page 2017

  4. [4]

    Towards better understanding of gradient-based attribution methods for Deep Neural Networks

    Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. Towards better under- standing of gradient-based attribution methods for deep neural networks.arXiv preprint arXiv:1711.06104, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  5. [5]

    right to explanation

    Bryce Goodman and Seth Flaxman. European union regulations on algorithmic decision-making and a “right to explanation”.AI Magazine, 38(3):50–57, 2017

    work page 2017

  6. [6]

    Debugging machine learning models

    Gabriel Cadamuro, Ran Gilad-Bachrach, and Xiaojin Zhu. Debugging machine learning models. InICML Workshop on Reliable Machine Learning in the Wild, volume 103, 2016

    work page 2016

  7. [7]

    2020 , journal =

    Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations.arXiv preprint arXiv:2011.05429, 2020

    work page arXiv 2011

  8. [8]

    Learning important features through propagating activation differences

    Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. InInternational Conference on Machine Learning, pages 3145–3153. PMLR, 2017

    work page 2017

  9. [9]

    Post hoc explanations may be ineffective for detecting unknown spurious correlation

    Julius Adebayo, Michael Muelly, Harold Abelson, and Been Kim. Post hoc explanations may be ineffective for detecting unknown spurious correlation. InInternational Conference on Learning Representations, 2022

    work page 2022

  10. [10]

    Deep learning-based electroencephalography analysis: a systematic review

    Yannick Roy, Hubert Banville, Isabela Albuquerque, Alexandre Gramfort, Tiago H Falk, and Jocelyn Faubert. Deep learning-based electroencephalography analysis: a systematic review. Journal of Neural Engineering, 16(5):051001, 2019

    work page 2019

  11. [11]

    A survey on explainable artificial intelligence (xai): Toward medical xai.IEEE Transactions on Neural Networks and Learning Systems, 32(11):4793–4813, 2020

    Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (xai): Toward medical xai.IEEE Transactions on Neural Networks and Learning Systems, 32(11):4793–4813, 2020

    work page 2020

  12. [12]

    Matt: A manifold attention network for eeg decoding

    Yue-Ting Pan, Jing-Lun Chou, and Chun-Shu Wei. Matt: A manifold attention network for eeg decoding. InAdvances in Neural Information Processing Systems, volume 35, pages 31116–31129, 2022

    work page 2022

  13. [13]

    Impossibility theorems for feature attribution.Proceedings of the National Academy of Sciences, 121(2):e2304406120, 2024

    Blair Bilodeau, Natasha Jaques, Pang Wei Koh, and Been Kim. Impossibility theorems for feature attribution.Proceedings of the National Academy of Sciences, 121(2):e2304406120, 2024

    work page 2024

  14. [14]

    Do input gradients highlight discriminative features? InAdvances in Neural Information Processing Systems, volume 34, pages 2046–2059, 2021

    Harshay Shah, Prateek Jain, and Praneeth Netrapalli. Do input gradients highlight discriminative features? InAdvances in Neural Information Processing Systems, volume 34, pages 2046–2059, 2021

    work page 2046

  15. [15]

    On the (in)fidelity and sensitivity of explanations

    Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Suggala, David I Inouye, and Pradeep K Ravikumar. On the (in)fidelity and sensitivity of explanations. InAdvances in Neural Information Processing Systems, volume 32, 2019

    work page 2019

  16. [16]

    Evaluations and methods for explanation through robustness analysis.arXiv preprint arXiv:2006.00442, 2020

    Cheng-Yu Hsieh, Chih-Kuan Yeh, Xuanqing Liu, Pradeep Ravikumar, Seungyeon Kim, Sanjiv Kumar, and Cho-Jui Hsieh. Evaluations and methods for explanation through robustness analysis.arXiv preprint arXiv:2006.00442, 2020

    work page arXiv 2006

  17. [17]

    Evaluating post-hoc explanations for graph neural networks via robustness analysis

    Junfeng Fang, Wei Liu, Yuan Gao, Zemin Liu, An Zhang, Xiang Wang, and Xiangnan He. Evaluating post-hoc explanations for graph neural networks via robustness analysis. InAdvances in Neural Information Processing Systems, volume 36, 2024. 10

    work page 2024

  18. [18]

    Axiomatic attribution for deep networks

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

    work page 2017

  19. [19]

    The (un)reliability of saliency methods

    Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un)reliability of saliency methods. InExplainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267–280. Springer, 2019

    work page 2019

  20. [20]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. InAdvances in Neural Information Processing Systems, volume 31, 2018

    work page 2018

  21. [21]

    An empirical comparison of deep learning explainability approaches for eeg using simulated ground truth.Scientific Reports, 13(1):17709, 2023

    Akshay Sujatha Ravindran and Jose Contreras-Vidal. An empirical comparison of deep learning explainability approaches for eeg using simulated ground truth.Scientific Reports, 13(1):17709, 2023

    work page 2023

  22. [22]

    Real time image saliency for black box classifiers

    Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. InAdvances in Neural Information Processing Systems, volume 30, 2017

    work page 2017

  23. [23]

    Evaluating feature importance estimates.arXiv preprint arXiv:1806.10758, 2018

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. Evaluating feature importance estimates.arXiv preprint arXiv:1806.10758, 2018

    work page arXiv 2018

  24. [24]

    A consis- tent and efficient evaluation strategy for attribution methods.arXiv preprint arXiv:2202.00449, 2022

    Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consis- tent and efficient evaluation strategy for attribution methods.arXiv preprint arXiv:2202.00449, 2022

    work page arXiv 2022

  25. [25]

    Sanity checks for saliency metrics

    Richard Tomsett, Dan Harborne, Supriyo Chakraborty, Prudhvi Gurram, and Alun Preece. Sanity checks for saliency metrics. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6021–6029, 2020

    work page 2020

  26. [26]

    Fidelity of interpretability methods and perturba- tion artifacts in neural networks.arXiv preprint arXiv:2203.02928, 2022

    Lennart Brocki and Neo Christopher Chung. Fidelity of interpretability methods and perturba- tion artifacts in neural networks.arXiv preprint arXiv:2203.02928, 2022

    work page arXiv 2022

  27. [27]

    Geometric remove-and-retrain (goar): Coordinate-invariant explainable ai assessment

    Yong-Hyun Park, Junghoon Seo, Bomseok Park, Seongsu Lee, and Junghyo Jo. Geometric remove-and-retrain (goar): Coordinate-invariant explainable ai assessment. InXAI in Action: Past, Present, and Future Applications. Springer, 2023

    work page 2023

  28. [28]

    Evaluation of post-hoc interpretability methods in time-series classification.Nature Machine Intelligence, 5 (3):250–260, 2023

    Hugues Turbé, Mina Bjelogrlic, Christian Lovis, and Gianmarco Mengaldo. Evaluation of post-hoc interpretability methods in time-series classification.Nature Machine Intelligence, 5 (3):250–260, 2023

    work page 2023

  29. [29]

    Explaining and Harnessing Adversarial Examples

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversar- ial examples.arXiv preprint arXiv:1412.6572, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  30. [30]

    On the vulnerability of cnn classifiers in eeg-based bcis.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(5):814–825, 2019

    Xiao Zhang and Dongrui Wu. On the vulnerability of cnn classifiers in eeg-based bcis.IEEE Transactions on Neural Systems and Rehabilitation Engineering, 27(5):814–825, 2019

    work page 2019

  31. [31]

    Reliable evaluation of attribution maps in cnns: A perturbation-based approach.International Journal of Computer Vision, 133: 2392–2409, 2025

    Laura Nieradzik, Heike Stephani, and Jan Keuper. Reliable evaluation of attribution maps in cnns: A perturbation-based approach.International Journal of Computer Vision, 133: 2392–2409, 2025. doi: 10.1007/s11263-024-02282-6

    work page doi:10.1007/s11263-024-02282-6 2025

  32. [32]

    Toward the applica- tion of xai methods in eeg-based systems.arXiv preprint arXiv:2210.06554, 2022

    Andrea Apicella, Francesco Isgrò, Andrea Pollastro, and Roberto Prevete. Toward the applica- tion of xai methods in eeg-based systems.arXiv preprint arXiv:2210.06554, 2022

    work page arXiv 2022

  33. [33]

    Towards best practice of interpreting deep learning models for eeg-based brain computer interfaces.Frontiers in Computational Neuroscience, 17:1232925, 2023

    Jian Cui, Liqiang Yuan, Zhaoxiang Wang, Ruilin Li, and Tianzi Jiang. Towards best practice of interpreting deep learning models for eeg-based brain computer interfaces.Frontiers in Computational Neuroscience, 17:1232925, 2023

    work page 2023

  34. [34]

    Evaluation of interpretability for deep learning algorithms in eeg emotion recognition: A case study in autism.Artificial Intelligence in Medicine, 143:102545, 2023

    Juan Manuel Mayor Torres, Sara Medina-DeVilliers, Tessa Clarkson, Matthew D Lerner, and Giuseppe Riccardi. Evaluation of interpretability for deep learning algorithms in eeg emotion recognition: A case study in autism.Artificial Intelligence in Medicine, 143:102545, 2023

    work page 2023

  35. [35]

    Journal of the Franklin Institute , year =

    Sören Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian La- puschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark.Journal of the Franklin Institute, 361(1):418–428, 2024. doi: 10.1016/j.jfranklin.2023.11.038. 11

    work page doi:10.1016/j.jfranklin.2023.11.038 2024

  36. [36]

    Metrics for saliency map evaluation of deep learning explanation methods, 2022

    Thibault Gomez, Thomas Fréour, and Harold Mouchère. Metrics for saliency map evaluation of deep learning explanation methods, 2022

    work page 2022

  37. [37]

    A comparative evaluation of explainability techniques for image data.Scientific Reports, 15:41898, 2025

    Mykyta Skliarov, Radwa El Shawi, Chedia Dhaoui, et al. A comparative evaluation of explainability techniques for image data.Scientific Reports, 15:41898, 2025. doi: 10.1038/s41598-025-25839-y

    work page doi:10.1038/s41598-025-25839-y 2025

  38. [38]

    F-fidelity: A robust framework for faithfulness evaluation of explainable ai

    Xu Zheng, Farhad Shirani, Zhuomin Chen, Chaohao Lin, Wei Cheng, Wenbo Guo, and Dong- sheng Luo. F-fidelity: A robust framework for faithfulness evaluation of explainable ai. In International Conference on Learning Representations, 2025. Poster

    work page 2025

  39. [39]

    To- wards bridging the gap between computational intelligence and neuroscience in brain-computer interfaces.Frontiers in Neuroinformatics, 15:699840, 2021

    Avinash Kumar Singh, Guillermo Sahonero-Alvarez, Mufti Mahmud, and Luigi Bianchi. To- wards bridging the gap between computational intelligence and neuroscience in brain-computer interfaces.Frontiers in Neuroinformatics, 15:699840, 2021

    work page 2021

  40. [40]

    Explainable artificial intelligence approaches for brain-computer interfaces: A review and design space.Journal of Neural Engineering, 2024

    Param Rajpura, Hubert Cecotti, and Yogesh Kumar Meena. Explainable artificial intelligence approaches for brain-computer interfaces: A review and design space.Journal of Neural Engineering, 2024

    work page 2024

  41. [41]

    Towards Deep Learning Models Resistant to Adversarial Attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  42. [42]

    A benchmark for inter- pretability methods in deep neural networks

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for inter- pretability methods in deep neural networks. InNeurIPS, 2019

    work page 2019

  43. [43]

    Towards robust evaluation of explainable artificial intelligence methods

    Andrea Apicella et al. Towards robust evaluation of explainable artificial intelligence methods. Pattern Recognition Letters, 2022

    work page 2022

  44. [44]

    On the evaluation of saliency methods

    Lukasz Brocki and Ngan Chung. On the evaluation of saliency methods. InICLR Workshop, 2022

    work page 2022

  45. [45]

    Explaining deep learning models: A comprehensive survey.IEEE Transac- tions, 2023

    Zhiyong Cui et al. Explaining deep learning models: A comprehensive survey.IEEE Transac- tions, 2023

    work page 2023

  46. [46]

    Bci competition 2008–graz data set a.Institute for Knowledge Discovery (Laboratory of Brain-Computer Interfaces), Graz University of Technology, 2008

    Clemens Brunner, Robert Leeb, Gernot R Müller-Putz, Alois Schlögl, and Gert Pfurtscheller. Bci competition 2008–graz data set a.Institute for Knowledge Discovery (Laboratory of Brain-Computer Interfaces), Graz University of Technology, 2008

    work page 2008

  47. [47]

    Bci challenge: Event-related negativity dataset, 2014

    Jérémie Mattout and K Kan. Bci challenge: Event-related negativity dataset, 2014. URL https://www.kaggle.com/c/inria-bci-challenge. Kaggle dataset

    work page 2014

  48. [48]

    A multi-modal dataset for steady-state visual evoked potential-based brain-computer interfaces.MAMEM SSVEP Dataset, 2007

    Victor Martinez-Cagigal, Enrique Santamaria-Vazquez, and Roberto Hornero. A multi-modal dataset for steady-state visual evoked potential-based brain-computer interfaces.MAMEM SSVEP Dataset, 2007

    work page 2007

  49. [49]

    Karol J. Piczak. Esc: Dataset for environmental sound classification. InACM Multimedia, 2015

    work page 2015

  50. [50]

    Kroos, O

    C. Kroos, O. Bones, Y . Cao, L. Harris, P. J. Jackson, W. J. Davies, W. Wang, T. J. Cox, and M. D. Plumbley. Generalisation in environmental sound classification: The making sense of sounds dataset. InICASSP, pages 8082–8086, 2019

    work page 2019

  51. [51]

    Parkhi, Andrea Vedaldi, and Andrew Zisserman

    Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Cats and dogs. InCVPR, 2012

    work page 2012

  52. [52]

    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. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009

    work page 2009

  53. [53]

    Bhuvaji, A

    S. Bhuvaji, A. Kadam, P. Bhumkar, and S. Dedge. Brain tumor classification (mri). Kaggle, 2020

    work page 2020

  54. [54]

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

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. InarXiv preprint arXiv:1312.6034, 2013. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  55. [55]

    SmoothGrad: removing noise by adding noise

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smooth- grad: removing noise by adding noise. InarXiv preprint arXiv:1706.03825, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  56. [56]

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PLOS ONE, 2015

    Sebastian Bach et al. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation.PLOS ONE, 2015

    work page 2015

  57. [57]

    Towards better understanding of gradient-based attribution methods

    Marco Ancona et al. Towards better understanding of gradient-based attribution methods. In ICLR, 2018

    work page 2018

  58. [58]

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

    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. InICCV, 2017

    work page 2017

  59. [59]

    Grad-cam++: Improved visual explanations for deep convolutional networks

    Aditya Chattopadhyay et al. Grad-cam++: Improved visual explanations for deep convolutional networks. InWACV, 2018

    work page 2018

  60. [60]

    Score-cam: Score-weighted visual explanations for convolutional neural net- works

    Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural net- works. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 111–119, 2020. doi: 10.1109/CVPRW50498.2020.00020

    work page doi:10.1109/cvprw50498.2020.00020 2020

  61. [61]

    Smooth grad- cam++: An enhanced inference level visualization technique for deep convolutional neural network models.arXiv preprint arXiv:1908.01224, 2019

    Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. Smooth grad- cam++: An enhanced inference level visualization technique for deep convolutional neural network models.arXiv preprint arXiv:1908.01224, 2019

    work page arXiv 1908

  62. [62]

    Gert Pfurtscheller, Clemens Brunner, Alois Schlögl, and F. H. Lopes da Silva. Mu rhythm (de)synchronization and eeg single-trial classification of different motor imagery tasks.Neu- roImage, 31(1):153–159, 2006

    work page 2006

  63. [63]

    TIMING: Temporality-aware integrated gradients for time series explanation

    Hyeongwon Jang, Changhun Kim, and Eunho Yang. TIMING: Temporality-aware integrated gradients for time series explanation. InForty-second International Conference on Machine Learning, 2025. URLhttps://openreview.net/forum?id=qOgKMqv9T7

    work page 2025

  64. [64]

    Robust learning from corrupted eeg with self-supervised learning.NeuroImage, 251: 118994, 2022

    Hubert Banville, Omar Chehab, Aapo Hyvarinen, Denis-Alexandre Engemann, and Alexandre Gramfort. Robust learning from corrupted eeg with self-supervised learning.NeuroImage, 251: 118994, 2022

    work page 2022

  65. [65]

    Stephan Leske and Sarang S. Dalal. Reducing power line noise in eeg and meg data via spectrum interpolation.NeuroImage, 189:763–776, 2019

    work page 2019

  66. [66]

    Biyu J. He. Scale-free brain activity: past, present, and future.Trends in Cognitive Sciences, 18 (9):480–487, 2014

    work page 2014

  67. [67]

    Parameterizing neural power spectra into periodic and aperiodic components.Nature Neuroscience, 23(12): 1655–1665, 2020

    Thomas Donoghue, Matthijs Haller, Erik J Peterson, Paroma Varma, Padraig Sebastian, Ruijie Gao, Takashi Noto, Antonio H Lara, Jonathan D Wallis, Robert T Knight, et al. Parameterizing neural power spectra into periodic and aperiodic components.Nature Neuroscience, 23(12): 1655–1665, 2020

    work page 2020

  68. [68]

    Multipoint fractional brownian bridge: construction and applications

    Tobias Friedrich and et al. Multipoint fractional brownian bridge: construction and applications. Stochastic Processes and their Applications, 2020

    work page 2020

  69. [69]

    Springer, 2013

    Jan Beran.Long-memory processes: probabilistic properties and statistical methods. Springer, 2013

    work page 2013

  70. [70]

    Kannathal and et al

    N. Kannathal and et al. Analysis of eeg signals using fractal dimension.Biomedical Signal Processing and Control, 2005

    work page 2005

  71. [71]

    Mandelbrot and John W

    Benoit B. Mandelbrot and John W. Van Ness. Fractional brownian motions, fractional noises and applications.SIAM Review, 10(4):422–437, 1968

    work page 1968

  72. [72]

    Davies and D

    Robert B. Davies and D. S. Harte. Tests for hurst effect.Biometrika, 74(1):95–101, 1987

    work page 1987

  73. [73]

    Simulation of fractional brownian motion.Master’s thesis, University of Twente, 2004

    Ton Dieker. Simulation of fractional brownian motion.Master’s thesis, University of Twente, 2004. 13

    work page 2004

  74. [74]

    Lawhern and et al

    Vernon J. Lawhern and et al. Eegnet: A compact convolutional neural network for eeg-based bcis.Journal of Neural Engineering, 2018

    work page 2018

  75. [75]

    Spatial component-wise convolutional network (sccnet) for motor-imagery eeg classification

    Chun-Shu Wei, Toshiaki Koike-Akino, and Ye Wang. Spatial component-wise convolutional network (sccnet) for motor-imagery eeg classification. InProceedings of the 9th International IEEE/EMBS Conference on Neural Engineering (NER), pages 328–331. IEEE, 2019

    work page 2019

  76. [76]

    Interpretable cnn for eeg.IEEE TBME, 2022

    Zhiyuan Cui and et al. Interpretable cnn for eeg.IEEE TBME, 2022. A Appendix A.1 Compute Resources All experiments were conducted on GPU-based workstations. The main experiments were run using NVIDIA RTX 4090 GPUs with CUDA acceleration. For each dataset–model–masking configuration, we generated post-hoc saliency maps, constructed MoRF and LeRF masking or...

    work page 2022