Pith. sign in

REVIEW 4 major objections 6 minor 63 references

EVolutionary Independent DEtermiNistiC Explanation

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read EVIDENCE claims to extract, deterministically and without model access, the exact signals a frozen black-box classifier relies on, with a proof that its filtered output converges.

desk verdict Frequency-domain RISE variant with a leaky test protocol and an invalid convergence proof; the application is real but the central claims don't hold. read the letter →

arxiv 2501.16357 v1 pith:4UAWB5S4 submitted 2025-01-20 cs.LG cs.AIeess.SP

Vincenzo Dentamaro , Paolo Giglio , Donato Impedovo , Giuseppe Pirlo This is my paper · ORCID
classification cs.LGcs.AIeess.SP
keywords explainableartificialintelligencedeterministicexplanationmodel-agnosticspectrogramfilteringconvergenceproofaudioclassificationCOVID-19diagnosisblack-boxmodels
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 introduces EVIDENCE, a method intended to give deterministic, model-agnostic explanations for black-box classifiers, aimed primarily at audio and other time-varying signals. The central claim is that by generating many random partial copies of an input, scoring each copy by how well the frozen model agrees with the known true label, and averaging the best-scoring copies, the result converges to a filtered version of the input that contains exactly the features the model uses for its decision. If correct, this would mean explanations are reproducible, do not depend on model internals, and can be used to improve classification by removing noise. The paper reports that feeding such filtered spectrograms to a frozen residual network improves COVID-19 detection precision by 32% and AUC by 16% relative to baseline, and outperforms the other explainability methods tested.

What carries the argument

The central object is the EVIDENCE filter, the matrix $\chi = \lim_{n\to\infty} \frac{1}{n}\sum_{c=1}^n h'_c \mathbf{Q}_c$, where each $\mathbf{Q}_c = \mathbf{M} \circ \mathbf{F}_c$ is the input spectrogram multiplied by a binary frequency-chunk mask and $h'_c$ weights copies with low cross-entropy against the known label. The proof's load-bearing identity is equation (13): the limit equals $h'_c \mathbf{A}' \circ \mathbf{M}$, with $\mathbf{A}'$ recording the frequency with which each spectrogram cell survives in the selected population. That identity turns a population of random masked copies into a deterministic frequency-importance map and a filter mask in one operation.

What would settle it

Build EVIDENCE filters from training-set labels only, then apply them to the held-out test items; if the AUC gain over the unfiltered baseline disappears or reverses, the reported 16% gain is attributable to test-label access rather than to extraction of reusable decision-relevant signals.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the EVIDENCE limit exists and equals a weighted input: for each entry $(i,j)$ of the input matrix $\mathbf{M}$, the weighted average of the selected partial inputs $\mathbf{Q}_c$ satisfies $\lim_{n\to\infty} \frac{1}{n}\sum_{c=1}^{n} h'_c \mathbf{Q}^c_{i,j} = h'_c A'_{i,j} \mathbf{M}_{i,j}$, where $h'_c = 1/(h_c+1)$ is the inverse cross-entropy weight and $A'_{i,j}$ is the fraction of surviving partial inputs that keep entry $(i,j)$ nonzero. The proof proceeds by bounding the average between $h'_c\min(\mathbf{M})$ and $h'_c\max(\mathbf{M})$ and then applying a frequency-count argument with the law of large numbers. The paper interprets this limit as a likelihood-weight matrix: each input coefficient survives with a weight equal to how often it appears in low-cross-entropy partial inputs, so the output $\chi$ is at once an explanation and a filter. Empirically, the paper claims that feeding the filtered spectrograms to a frozen residual network raises the macro F1 on COVID-19 audio from 0.65 to 0.94, reaches 0.997 macro F1 on Parkinson voice data, and holds AUC 0.996 on GTZAN, surpassing the other explainability methods compared.

Load-bearing premise

The method assumes it is valid to measure explanation quality by using each test item's known true label to select the partial inputs that form that item's filter, and then reporting the frozen model's accuracy on the filtered test set.

Editorial extensions

If this is right

  • Filtering an input with the EVIDENCE mask and re-running the frozen model should preserve or improve accuracy whenever the model's decision depends on a reproducible subset of frequencies.
  • Because the output depends only on the input, the frozen model, and the task, repeated runs on the same input give identical explanations, unlike stochastic perturbation methods.
  • The weight matrix $\mathbf{A}'$ doubles as a frequency-importance map, so EVIDENCE can be used to identify which spectral bands drive a model's decision.
  • Since the algorithm only consumes the model's output scores, it should apply to any frozen classifier without architectural access, allowing direct comparison of what different models rely on.
  • The convergence argument is intended to extend from frequency-chunk masking on audio to n-dimensional inputs such as images and video.

Reading between the lines

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

  • Editorial inference: the reported accuracy gains should be re-tested with filters built from training-set labels and applied to held-out test items; if the gains vanish, the headline numbers would reflect label access rather than extraction of decision-relevant features.
  • Editorial inference: the proof establishes that the weighted average of masked inputs converges, not that the surviving bands cause the model's decision; causal claims would require intervention experiments that remove the selected bands and observe decision changes.
  • Editorial inference: for the practical 5000-iteration implementation, the output depends on the random sampling unless the seed is fixed, so the strict determinism claim belongs to the infinite limit rather than the finite run.
  • Editorial inference: applying EVIDENCE to 2D images with two-dimensional masks would make it directly comparable to region-based saliency methods and can be tested on standard image classification benchmarks.
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 / 6 minor

Summary. The paper introduces EVIDENCE, an XAI method that generates many random binary masks over a Mel spectrogram, scores each masked spectrogram by cross-entropy against the true class label, retains the lowest-cross-entropy subset, and averages them to produce an importance-weighted filter. It claims a mathematical proof (Section II.C, Eq. 13) that this average converges to h'_c A'_ij M_ij, and reports experiments on COVID-19 audio, Parkinson's voice recordings, and GTZAN in which filtering with EVIDENCE improves a frozen ResNet50 relative to LIME, SHAP Deep Explainer, and GradCAM.

Significance. The intended contribution is potentially valuable: a deterministic, model-agnostic method for extracting decision-relevant spectrogram regions, with empirical comparisons to three standard XAI baselines on three datasets. The Appendix's pseudocode makes the procedure concrete. However, the two pillars of the paper have load-bearing problems: the convergence proof factors out a per-sample quantity without justification and asserts convergence of an empirical frequency without a sampling law, while the experimental pipeline uses the true test label to construct the filter and then scores the same test items. These issues invalidate the central claims as stated, so the significance of the manuscript in its current form is not established.

major comments (4)
  1. [Section II.C, Eq. (9)] The proof factors h'_c out of the average, replacing lim_{n→∞} (1/n) Σ_{c=1}^n h'_c max(Q_c) with h'_c max(M). This is not justified because h'_c = 1/(h_c+1) varies with c through the cross-entropy h_c of each selected masked spectrogram Q_c, and the selected set Q itself depends on n and on the threshold W. The same unjustified pull-out is repeated in Eqs. (11)-(12). Without an additional argument that h'_c is constant or that the weighted average converges to a constant times max(M), Eq. (9) does not establish the claimed convergence.
  2. [Section II.C, Eqs. (10)-(13)] The quantity A'_{ij} = N_n/n is an empirical frequency of selected masks in which Q_c^{ij} ≠ 0. The claim that A'_{ij} is independent of n because ψ is independent of n does not follow: for the random mask generation in Appendix A, N_n is random and n-dependent, and no sampling law is specified. In addition, the mask universe D'_{2,m} has cardinality 2^m and the selected set Q has size n ≤ 2^m, so with fixed m there is no infinite sequence over which n → ∞; the appeal to the law of large numbers is therefore not supported. Consequently Eq. (13), the central deterministic-convergence result, is not proved.
  3. [Section II.D and Appendix A, Algorithm 1] The filter construction uses one_hot_instance_label—the true label of the test example—as input, computes categorical cross-entropy between the model's output on each random mask and this true label, keeps the samples with lowest cross-entropy, and uses them to zero out parts of the same test spectrogram. The frozen model is then scored on these filtered test items. This is label leakage: the reported improvements in Table 4 are not independent evidence that EVIDENCE identifies model-relevant features, since the filter has been selected using the very labels being predicted. A valid evaluation must build filters without test labels and then evaluate on held-out examples.
  4. [Section III, Table 4] The headline numbers are reported without confidence intervals or variance estimates on small test sets (32 COVID-19 items, 46 Parkinson items, and 100 GTZAN items), despite the stochastic mask sampling in EVIDENCE. The abstract's '16% AUC gain' also does not match Table 4, where COVID-19 AUC is 0.82 at baseline and 0.99 with EVIDENCE (a 0.17 absolute increase); the reported gains should be recomputed and stated precisely, with error bars.
minor comments (6)
  1. [Section headings] The manuscript contains a section-numbering error: 'II. METHODS' appears after 'II. STATE OF THE ART'; the methods section should be renumbered.
  2. [Section II.C, Thesis statement] The notation ∋′ and the phrase '∋′ h_c ≤ W' are nonstandard and undefined; please define them.
  3. [Equation (4)] Equation (4) has an indexing ambiguity: the condition '1 ≤ z ≤ m' is redundant with the definition of q, and the interval [z·m0, (z+1)·m0] should be stated relative to zero-based or one-based indexing.
  4. [Appendix A, Algorithm 1] The pseudocode includes unused parameters (Y_train, instance_label) and uses random selection in PARALLEL; if determinism is claimed, specify that results are reproducible only with a fixed random seed.
  5. [Abstract and Conclusions] Percentage gains are reported without specifying absolute vs. relative; for example, '32% improvement in precision' corresponds to 0.32 absolute percentage points in Table 4.
  6. [Section III, EVIDENCE configuration] The threshold W and filter fraction are not reported for each dataset, although they are free parameters; give the values and a sensitivity analysis.

Circularity Check

2 steps flagged · score 8.0 of 10

EVIDENCE's central 'convergence' theorem is a restatement of its own definition of A′, and the reported gains are obtained by filtering with test labels before scoring on the same labels.

  1. self definitional [Section II.B and Appendix A, Algorithm 1]
    "lim_(n->∞) (1/n) Σ_{c=1}^n h′_c · Q_c^{i,j} = h′_c · A′_{i,j} · M_{i,j} ... where A′_{i,j} represents a fraction of the overall set Q of functions Q_c that satisfy Q_c^{i,j} ≠ 0 for a given threshold W and a given model ψ."

    By Eq. (10), Σ_c Q_c^{i,j} = N_n M_{i,j} + R_n·0, and the paper defines A′_{i,j} = N_n/n via the line N_n = A′_{i,j}·n. Substituting into Eq. (12) turns the left-hand average into h′_c (N_n/n) M_{i,j}, which is exactly h′_c A′_{i,j} M_{i,j}. Equation (13) is therefore the definition of the empirical frequency of mask coverage, not a derived limit. The 'convergence' claim reduces to the assertion that this empirical frequency has a limit; no sampling law or model-independent argument supplies that limit. The pull-out of h′_c from the sum is also not justified unless all selected partial spectrograms share one cross-entropy weight, which the paper does not establish.

  2. fitted input called prediction [Section II.B and Appendix A, Algorithm 1]
    "the EVIDENCE algorithm pipeline starts with the AI model directly applied on a signal to be classified whose ground truth value is already known ... the cross-entropy operator is applied between the model output and the ground truth value of the signal class; Algorithm 1: function EVIDENCE(sound, model, one_hot_instance_label, ...) ... cce = calculate categorical cross entropy between one_hots and probas ... scores[cce[i]] = samples[i]."

    The filter for each test item is built by scoring random partial spectrograms against the true one-hot test label, keeping only the lowest-cross-entropy partial spectrograms, and averaging them. The same labeled test items are then classified by the frozen ResNet50 to produce the reported precision and AUC gains. Thus the selection criterion is agreement with the ground-truth label, and the evaluation metric is agreement with the same ground-truth label. The headline 32% precision and 16% AUC improvements are not independent evidence that EVIDENCE finds the model's decision-relevant features; they are forced by the label-based fitting protocol used to construct the filtered inputs.

full rationale

The mathematical core of EVIDENCE (Section II.C) defines A′_{i,j} as the fraction of selected masks that cover cell (i,j). Equation (13) then says the weighted average of the selected partial spectrograms equals h′_c times that fraction times M_{i,j}. This is a definition-level identity: for binary masks, the average of values M_{i,j} and 0 is exactly the frequency of 1 times M_{i,j}. The 'proof of convergence' therefore does not derive a model-independent signal; it merely names the empirical selection frequency and asserts it has a limit. The proof also assumes h′_c is constant across selected items, which is not shown and would not follow from the definitions. The experimental protocol is circular for the same reason: Algorithm 1 takes the true one-hot label as an input, uses it to compute cross-entropy, selects the partial spectrograms that minimize that cross-entropy, averages them into a filter, and then the frozen model is scored on the filtered test set. Reported gains reflect how well the inputs were selected to match the labels, not an independent test of explanation quality. Comparisons to LIME, SHAP, and GradCAM do not repair this because those methods are run through a different filtering procedure, and the EVIDENCE filter is constructed with label access. The self-citation in reference [62] is not load-bearing for the EVIDENCE derivation, so it does not add circularity. Overall, both the formal theorem and the headline empirical claims reduce by construction to the paper's own selection frequencies and label-based fitting, warranting a score of 8.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central algorithm depends on hand-set hyperparameters and unproven assumptions. The proof mostly restates the empirical frequency of selected masks, and the experiments use ground-truth labels during filter construction, which inflates the reported accuracy gains.

free parameters (5)
  • threshold W = 0.25 in Algorithm 1; W undefined in Eq. (7)
    Selects the subset Q of partial spectrograms; no sensitivity analysis is provided.
  • chunk width (chunks_hz) = 1-2 Hz depending on dataset
    Controls granularity of frequency bands; set manually per dataset.
  • features per mask = 45 (COVID/Parkinson), 200 (GTZAN)
    Number of random frequency chunks kept in each partial spectrogram.
  • iterations K = 500 (GTZAN), 5000 (COVID/Parkinson)
    Number of random masks generated; directly affects runtime and the approximation of the claimed limit.
  • filter fraction = 30% (text)
    Fraction of original spectrogram information retained in each partial version; stated as a stochastic filter parameter.
assumptions (5)
  • domain assumption Availability of ground truth for every input being explained
    EVIDENCE selects partial spectrograms with lowest cross-entropy to the true label; Section II.A states this assumption.
  • domain assumption The selected subset of partial spectrograms converges to a single condensed information content
    Section II.A assumes convergence of the population; used to justify the averaged mask as an explanation.
  • ad hoc to paper Model psi is independent of the cardinality n of Q and of D'_{2,m}
    Assumed in Section II.C to make A_ij independent of n; not established for finite samples.
  • ad hoc to paper h'_c can be treated as a constant in the sum over c
    Equation (11)-(12) pull h'_c out of the summation although h'_c = 1/(h_c+1) varies with c; this is an algebraic gap in the proof.
  • standard math Law of large numbers justifies A'_ij as an expected frequency
    Used in the corollary to interpret A'_ij as a frequency distribution; standard but invoked without formal conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EVolutionary Independent DEtermiNistiC Explanation." pith.science (2026). https://pith.science/paper/4UAWB5S4

@misc{pith2026250116357,
  author       = {Pith},
  title        = {Pith review of: EVolutionary Independent DEtermiNistiC Explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UAWB5S4}},
  note         = {Machine review of arXiv:2501.16357}
}
read the original abstract

The widespread use of artificial intelligence deep neural networks in fields such as medicine and engineering necessitates understanding their decision-making processes. Current explainability methods often produce inconsistent results and struggle to highlight essential signals influencing model inferences. This paper introduces the Evolutionary Independent Deterministic Explanation (EVIDENCE) theory, a novel approach offering a deterministic, model-independent method for extracting significant signals from black-box models. EVIDENCE theory, grounded in robust mathematical formalization, is validated through empirical tests on diverse datasets, including COVID-19 audio diagnostics, Parkinson's disease voice recordings, and the George Tzanetakis music classification dataset (GTZAN). Practical applications of EVIDENCE include improving diagnostic accuracy in healthcare and enhancing audio signal analysis. For instance, in the COVID-19 use case, EVIDENCE-filtered spectrograms fed into a frozen Residual Network with 50 layers improved precision by 32% for positive cases and increased the area under the curve (AUC) by 16% compared to baseline models. For Parkinson's disease classification, EVIDENCE achieved near-perfect precision and sensitivity, with a macro average F1-Score of 0.997. In the GTZAN, EVIDENCE maintained a high AUC of 0.996, demonstrating its efficacy in filtering relevant features for accurate genre classification. EVIDENCE outperformed other Explainable Artificial Intelligence (XAI) methods such as LIME, SHAP, and GradCAM in almost all metrics. These findings indicate that EVIDENCE not only improves classification accuracy but also provides a transparent and reproducible explanation mechanism, crucial for advancing the trustworthiness and applicability of AI systems in real-world settings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 55 canonical work pages

  1. [1]

    and Mongia, S., 2023

    Gupta, L.K., Koundal, D. and Mongia, S., 2023. Explainable Methods for Image -Based Deep Learning: A Review. Archives of Computational Methods in Engineering , pp.1-16

    work page 2023

  2. [2]

    and Bouridane, A., 2023

    Chaddad, A., Peng, J., Xu, J. and Bouridane, A., 2023. Survey of Explainable AI Techniques in Healthcare. Sensors, 23(2), p.634

    work page 2023

  3. [3]

    and Ranjan, R., 20 23

    Dwivedi, R., Dave, D., Naik, H., Singhal, S., Omer, R., Patel, P., Qian, B., Wen, Z., Shah, T., Morgan, G. and Ranjan, R., 20 23. Explainable AI (XAI): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9), pp.1-33

    work page

  4. [4]

    and Omlin, C., 2023

    Saeed, W. and Omlin, C., 2023. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, p.110273

    work page 2023

  5. [5]

    and Han, T.Y.J., 2022

    Zhong, X., Gallagher, B., Liu, S., Kailkhura, B., Hiszpanski, A. and Han, T.Y.J., 2022. Explainable machine learning in materials science. npj Computational Materials, 8(1), p.204

    work page 2022

  6. [6]

    and Stieglitz, S., 2022

    Hofeditz, L., Clausen, S., Rieß, A., Mirbabaie, M. and Stieglitz, S., 2022. Applying XAI to an AI-based system for candidate management to mitigate bias and discrimination in hiring. Electronic Markets, pp.1-27

    work page 2022

  7. [8]

    and Capuzzo, D., 2022

    Visani, G., Bagli, E., Chesani, F., Poluzzi, A. and Capuzzo, D., 2022. Statistical stability indices for LIME: Obtaining reliable explanations for machine learning models. Journal of the Operational Research Society, 73(1), pp.91-101

    work page 2022

  8. [9]

    and Havens, T.C., 2022, December

    Hamilton, N., Webb, A., Wilder, M., Hendrickson, B., Blanck, M., Nelson, E., Roemer, W. and Havens, T.C., 2022, December. Enhancing Visualization and Explainability of Computer Vision Models with Local Interpretable Model -Agnostic Explanations (LIME). In 2022 IEEE Symposium Series on Computational Intelligence (SSCI) (pp. 604-611). IEEE

    work page 2022

Show all 63 references
  1. [10]

    and Domingo-Ferrer, J., 2019

    Blanco-Justicia, A. and Domingo-Ferrer, J., 2019. Machine learning explainability through comprehensible decision trees. In Machine Learning and Knowledge Extraction: Third IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross -Domain Conference, CD-MAKE 2019, Canterbu...

    2019

  2. [11]

    and De Waal, A., 2020

    Derks, I.P. and De Waal, A., 2020. A taxonomy of explainable Bayesian networks. In Artificial Intelligence Research: First Southern African Conference for AI Research, SACAIR 2020, Muldersdrift, South Africa, February 22-26, 2021, Proceedings 1 (pp. 220-235). Springer Internat...

    2020

  3. [12]

    and Bielza, C., 2022, June

    Valero-Leal, E., Larranaga, P. and Bielza, C., 2022, June. Extending MAP -independence for Bayesian network explainability. In Proceedings of the First International Workshop on Heterodox Methods for Interpretable and Efficient AI

    2022

  4. [13]

    and Pereira -Fariña, M., 2021

    Stepin, I., Alonso, J.M., Catala, A. and Pereira -Fariña, M., 2021. A survey of contrastive and counterfactual explanation generation methods for explainable artificial intelligence. IEEE Access, 9, pp.11974-12001

    2021

  5. [14]

    and Wang, M.D., 2020, September

    Shi, W., Tong, L., Zhuang, Y., Zhu, Y. and Wang, M.D., 2020, September. Exam: an explainable attention-based model for covid-19 automatic diagnosis. In Proceedings of the 11th ACM international conference on bioinformatics, computational biology and health informatics (pp. 1-6)

    2020

  6. [15]

    and Rodriguez Martinez, M., 2019

    Manica, M., Oskooei, A., Born, J., Subramanian, V., Sáez-Rodríguez, J. and Rodriguez Martinez, M., 2019. Toward explainable anticancer compound sensitivity prediction via multimodal attention-based convolutional encoders. Molecular pharmaceutics, 16(12), pp.4797-4806

    2019

  7. [16]

    and Srivastava, G., 2021, July

    Ahmed, U., Lin, J.C.W. and Srivastava, G., 2021, July. Fuzzy explainable attention -based deep active learning on mental -health data. In 2021 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) (pp. 1-6). IEEE

    2021

  8. [17]

    and Balakrishnan, N., 2022

    Hariharan, S., Rejimol Robinson, R.R., Prasad, R.R., Thomas, C. and Balakrishnan, N., 2022. XAI for intrusion detection system: comparing explanations based on global and local scope. Journal of Computer Virology and Hacking Techniques, pp.1-23

    2022

  9. [18]

    and Rinzivillo, S., 2021

    Bodria, F., Giannotti, F., Guidotti, R., Naretto, F., Pedreschi, D. and Rinzivillo, S., 2021. Benchmarking and survey of explanation methods for black box models. arXiv preprint arXiv:2102.13076. 17 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE -CLICK HERE TO EDIT) <

    2021 arXiv

  10. [19]

    and Arens, M., 2021

    Buhrmester, V., Münch, D. and Arens, M., 2021. Analysis of explainers of black box deep neural networks for computer vision: A survey. Machine Learning and Knowledge Extraction, 3(4), pp.966-989

    2021

  11. [20]

    and Levron, Y., 2022

    Machlev, R., Heistrene, L., Perl, M., Levy, K.Y., Belikov, J., Mannor, S. and Levron, Y., 2022. Explainable Artificial Intell igence (XAI) techniques for energy and power systems: Review, challenges and opportunities. Energy and AI, p.100169

    2022

  12. [21]

    and Raychowdhury, A., 2022, October

    Bhat, A., Assoa, A.S. and Raychowdhury, A., 2022, October. Gradient Backpropagation based Feature Attribution to Enable Expla inable-AI on the Edge. In 2022 IFIP/IEEE 30th International Conference on Very Large Scale Integration (VLSI-SoC) (pp. 1-6). IEEE

    2022

  13. [22]

    and Viergever, M.A., 2022

    Van der Velden, B.H., Kuijf, H.J., Gilhuijs, K.G. and Viergever, M.A., 2022. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Medical Image Analysis, p.102470

    2022

  14. [23]

    and Rad, P., 2020

    Das, A. and Rad, P., 2020. Opportunities and challenges in explainable artificial intelligence (xai): A survey. arXiv preprint arXiv:2006.11371

    2020 arXiv

  15. [24]

    and Hamarneh, G., 2023

    Jin, W., Li, X., Fatehi, M. and Hamarneh, G., 2023. Generating Post -Hoc Explanation from Deep Neural Networks for Multi -Modal Medical Image Analysis Tasks. MethodsX, p.102009

    2023

  16. [25]

    and Chen, L., 2021, August

    Li, X.H., Shi, Y., Li, H., Bai, W., Cao, C.C. and Chen, L., 2021, August. An experimental study of quantitative evaluations on saliency methods. In Proceedings of the 27th ACM sigkdd conference on knowledge discovery & data mining (pp. 3200-3208)

    2021

  17. [26]

    and Sarkar, S., 2022

    Druc, S., Balu, A., Wooldridge, P., Krishnamurthy, A. and Sarkar, S., 2022. Concept Activation Vectors for Generating User -Defined 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 2993-3000)

    2022

  18. [27]

    and Lee, W., 2022

    Byun, S.Y. and Lee, W., 2022. Recipro-CAM: Gradient-free reciprocal class activation map. arXiv preprint arXiv:2209.14074

    2022 arXiv

  19. [28]

    and Kim, K.J., 2019, August

    Joo, H.T. and Kim, K.J., 2019, August. Visualization of deep reinforcement learning using grad -CAM: how AI plays atari games?. In 2019 IEEE Conference on Games (CoG) (pp. 1-2). IEEE

    2019

  20. [29]

    and Kose, U., 2022

    Marmolejo-Saucedo, J.A. and Kose, U., 2022. Numerical grad-CAM based explainable convolutional neural network for brain tumor diagnosis. Mobile Networks and Applications, pp.1-10

    2022

  21. [30]

    and Nowosielski, J.M., 2020

    Gorski, L., Ramakrishna, S. and Nowosielski, J.M., 2020. Towards grad-cam based explainability in a legal text processing pipeline. arXiv preprint arXiv:2012.09603

    2020 arXiv

  22. [31]

    and Groumpos, P.P., 2022

    Apostolopoulos, I.D., Athanasoula, I., Tzani, M. and Groumpos, P.P., 2022. An Explainable Deep Learning Framework for Detecti ng and Localising Smoke and Fire Incidents: Evaluation of Grad-CAM++ and LIME. Machine Learning and Knowledge Extraction, 4(4), pp.1124-1135

    2022

  23. [32]

    and Lucas, M., 2022

    Lerma, M. and Lucas, M., 2022. Grad-CAM++ is Equivalent to Grad-CAM With Positive Gradients. arXiv preprint arXiv:2205.10838

    2022 arXiv

  24. [33]

    and Choi, H.J., 2021

    Jung, Y.J., Han, S.H. and Choi, H.J., 2021. Explaining cnn and rnn using selective layer-wise relevance propagation. IEEE Access, 9, pp.18670-18681

    2021

  25. [34]

    and Mckeever, S., 2021

    Ullah, I., Rios, A., Gala, V. and Mckeever, S., 2021. Explaining deep learning models for tabular data using layer -wise relevance propagation. Applied Sciences, 12(1), p.136

    2021

  26. [35]

    and Weldermariam, K., 2019

    Omeiza, D., Speakman, S., Cintas, C. and Weldermariam, K., 2019. Smooth grad -cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224

    2019 arXiv

  27. [36]

    QUANTIFYING TRUST IN DEEP LEARNING WITH OBJECTIVE EXPLAINABLE AI METHODS FOR ECG CLASSIFICATION (Doctoral dissertation)

    Siddiqui, M.K., 2022. QUANTIFYING TRUST IN DEEP LEARNING WITH OBJECTIVE EXPLAINABLE AI METHODS FOR ECG CLASSIFICATION (Doctoral dissertation)

    2022

  28. [37]

    and Khan, N., 2021

    Zafar, M.R. and Khan, N., 2021. Deterministic local interpretable model-agnostic explanations for stable explainability. Machine Learning and Knowledge Extraction, 3(3), pp.525-541

    2021

  29. [38]

    and Fan, W., 2020

    Shi, S., Zhang, X. and Fan, W., 2020. A modified perturbed sampling method for local interpretable model -agnostic explanation. arXiv preprint arXiv:2002.07434

    2020 arXiv

  30. [39]

    and Seebacher, D., 2022, February

    Schlegel, U., Vo , D.L., Keim, D.A. and Seebacher, D., 2022, February. Ts-mule: Local interpretable model -agnostic explanations for time series forecast models. In Machine Learning and Principles and Practice of Knowledge Discovery in Databases: International Workshops of ECM...

    2022

  31. [40]

    and Khan, N.M., 2019

    Zafar, M.R. and Khan, N.M., 2019. DLIME: A deterministic local interpretable model-agnostic explanations approach for computer-aided diagnosis systems. arXiv preprint arXiv:1906.10263

    2019 arXiv

  32. [41]

    and Demeester, T., 2022, November

    López, Y.A., Diez, H.R.G., Toledano -López, O.G., Hidalgo-Delgado, Y., Mannens, E. and Demeester, T., 2022, November. DLIME -Graphs: A DLIME Extension Based on Triple Embedding for Graphs. In Knowledge Graphs and Semantic Web: 4th Iberoamerican Conference and third Indo-Americ...

    2022

  33. [42]

    and Suciu, D., 2022

    Van den Broeck, G., Lykov, A., Schleich, M. and Suciu, D., 2022. On the tractability of SHAP explanations. Journal of Artificial Intelligence Research, 74, pp.851-886

    2022

  34. [43]

    and Aznarte, J.L., 2020

    García, M.V. and Aznarte, J.L., 2020. Shapley additive explanations for NO2 forecasting. Ecological Informatics, 56, p.101039

    2020

  35. [44]

    and Pereira, J.M., 2022

    Chou, Y.L., Hsieh, C., Moreira, C., Ouyang, C., Jorge, J. and Pereira, J.M., 2022. Benchmark Evaluation of Counterfactual Algorithm s for XAI: From a White Box to a Black Box. arXiv preprint arXiv:2203.02399

    2022 arXiv

  36. [45]

    Explaining the output of a black box model and a white box model: an illustrative comparison

    Joel, V., 2020. Explaining the output of a black box model and a white box model: an illustrative comparison

    2020

  37. [46]

    and Bonnet, G., 2020

    Brkan, M. and Bonnet, G., 2020. Legal and technical feasibility of the GDPR’s quest for explanation of algorithmic decisions: of black boxes, white boxes and fata morganas. European Journal of Risk Regulation, 11(1), pp.18-50

    2020

  38. [47]

    and Omicini, A., 2020

    Calegari, R., Ciatto, G. and Omicini, A., 2020. On the integration of symbolic and sub-symbolic techniques for XAI: A survey. Intelligenza Artificiale, 14(1), pp.7-32

    2020

  39. [48]

    and Zhou, B., 2023

    Li, B., Qi, P., Liu, B., Di, S., Liu, J., Pei, J., Yi, J. and Zhou, B., 2023. Trustworthy ai: From principles to practices. ACM Computing Surveys, 55(9), pp.1-46

    2023

  40. [49]

    and Lenca, P., 2021, August

    Kaadoud, I.C., Fahed, L. and Lenca, P., 2021, August. Explainable AI: a narrative review at the crossroad of Knowledge Discov ery, Knowledge Representation and Representation Learning. In MRC 2021: Twelfth International Workshop Modelling and Reasoning in Context (Vol. 2995, p...

    2021

  41. [50]

    and Herrera, F., 2020

    Charte, D., Charte, F., del Jesus, M.J. and Herrera, F., 2020. An analysis on the use of autoencoders for representation learning: Fundamentals, learning task case studies, explainability and challenges. Neurocomputing, 404, pp.93-107

    2020

  42. [51]

    and Li, Y., 2021, April

    Qian, K., Danilevsky, M., Katsis, Y., Kawas, B., Oduor, E., Popa, L. and Li, Y., 2021, April. XNLP: A living survey for XAI research in natural language processing. In 26th International Conference on Intelligent User Interfaces-Companion (pp. 78-80)

    2021

  43. [52]

    and Ghosh, J., 2020

    Jain, A., Ravula, M. and Ghosh, J., 2020. Biased models have biased explanations. arXiv preprint arXiv:2012.10986

    2020 arXiv

  44. [53]

    and Ravikumar, P.K., 2019

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

    2019

  45. [54]

    and Bischl, B., 2021, February

    Molnar, C., Casalicchio, G. and Bischl, B., 2021, February. Interpretable machine learning –a brief history, state -of-the-art and challenges. In ECML PKDD 2020 Workshops: Workshops of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD ...

    2021

  46. [55]

    and Terzija, V., 2020

    Wang, X., Ding, L., Ma, Z., Azizipanah -Abarghooee, R. and Terzija, V., 2020. Perturbation -based sensitivity analysis of slow coherency with variable power system inertia. IEEE Transactions on Power Systems, 36(2), pp.1121-1129

    2020

  47. [56]

    and Natarajan, B., 2020

    Jhala, K., Pradhan, P. and Natarajan, B., 2020. Perturbation -based diagnosis of false data injection attack using distributed energy resources. IEEE Transactions on Smart Grid, 12(2), pp.1589-1601

    2020

  48. [57]

    and Miri, M., 2019

    Rashki, M., Azhdary Moghaddam, M. and Miri, M., 2019. System‐level reliability sensitivity analysis by using weighted average simulation method. Quality and Reliability Engineering International, 35(6), pp.1826-1845

    2019

  49. [58]

    and Du, Q., 2020

    Xu, J. and Du, Q., 2020. Adversarial attacks on text classification models using layer ‐wise relevance propagation. International Journal of Intelligent Systems , 35(9), pp.1397-1415

    2020

  50. [59]

    and Le-Khac, N.A., 2021

    Kuppa, A. and Le-Khac, N.A., 2021. Adversarial xai methods in cybersecurity. IEEE transactions on information forensics and security, 16, pp.4924-4938

    2021

  51. [60]

    S-lime: Stabilized-lime for model explanation

    Zhou, Z., Hooker, G., & Wang, F., 2021. S-lime: Stabilized-lime for model explanation. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining (pp. 2429-2438)

    2021

  52. [61]

    & Mascolo, C

    Han, J., Brown, C., Chauhan, J., Grammenos, A., Hasthanasombat, A., Spathis, D., ... & Mascolo, C. , 2021. Exploring automatic COVID -19 diagnosis via voice and symptoms from crowdsourced data. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal P...

    2021

  53. [62]

    AUCO ResNet: An end-to-end network for Covid-19 pre-screening from cough and breath

    Dentamaro, V., Giglio, P., Impedovo, D., Moretti, L., & Pirlo, G., 2022. AUCO ResNet: An end-to-end network for Covid-19 pre-screening from cough and breath. Pattern Recognition, 127, 108656

    2022

  54. [63]

    Acoustic tracking of pitch, modal, and subharmonic vibra tions of vocal folds in Parkinson’s disease and Parkinsonism,

    J. Hlavnicka, R. Cmejla, J. Klempir, E. Ruzicka, and J. Rusz, 2019, “Acoustic tracking of pitch, modal, and subharmonic vibra tions of vocal folds in Parkinson’s disease and Parkinsonism,” IEEE Access, vol. 7, pp. 150339–150354, doi: 10.1109/ACCESS.2019.2945874

    2019

  55. [64]

    Deep learning techniques for music genre classification and feature importance

    Srivastava, K., 2022. Deep learning techniques for music genre classification and feature importance. https://doi.org/10.3622 7/techrxiv.21265965 18 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE -CLICK HERE TO EDIT) < 19 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.