Pith. sign in

REVIEW 6 major objections 5 minor 37 references

Explainable embeddings with Distance Explainer

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

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

Pith's one-line read Distance Explainer turns pairwise embedding distances into attribution maps by summing the masks at the extreme distance ranks.

desk verdict Plausible RISE-style explainer for pairwise embedding distances that fills a real gap, but the central selection heuristic is unvalidated and the quantitative support is thinner than the claims. read the letter →

arxiv 2505.15516 v3 pith:KOVJDIID submitted 2025-05-21 cs.LG cs.AIcs.CLcs.CV

Christiaan Meijer , E. G. Patrick Bos This is my paper · ORCID
classification cs.LGcs.AIcs.CLcs.CV
keywords explainableAIattributionmapsembeddedspacespairwisedistanceexplanationRISECLIPcosinerobustness
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

Distance Explainer is a post-hoc explanation method for pairwise distances in learned embedding spaces. It takes one input as the reference and one as the item to be explained, generates many random masks, embeds each masked version, and ranks the masked versions by cosine distance to the reference embedding. Instead of averaging all masks weighted by a score, it keeps only the masks at the extreme low and extreme high distance ranks and sums them into an attribution map, optionally subtracting the two extremes in a 'mirror' mode. The paper claims this identifies the image regions that push the two embedded points together or apart, and supports the claim with faithfulness, robustness, and model-randomization evaluations on image-image and image-caption pairs. A sympathetic reader would care because the method is modality-agnostic and addresses a gap: most XAI explains a single prediction, not a relationship between two points in a shared vector space.

What carries the argument

The load-bearing mechanism is distance-ranked mask filtering. Each random mask is scored not by a class probability but by the cosine distance $d_{\cos}(M_i(e;c), r)$ between the masked embedding and the reference embedding; masks are sorted by this distance, only the top and bottom fixed fractions are kept, and the kept masks are summed with the bottom set negated in mirror mode. This replaces RISE's weighted sum over all masks and guarantees a fixed percentage of effective masks, which matters because raw cosine-distance differences between high-dimensional embeddings are typically below $10^{-4}$, too small to use as weights. The mirror combination is the paper's device for canceling noise while producing a signed map whose red regions decrease distance to the reference and blue regions increase it.

What would settle it

Take two synthetic images that are identical except for a single known patch, use one as the reference, run Distance Explainer on the other, and check whether the attribution map concentrates on that patch; if the extreme-distance masks do not recover the known patch better than a random mask baseline, the central mechanism fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a mechanism for local explanation of distances in arbitrary embedded spaces. The method, called Distance Explainer, applies random binary masks to one of the two items, computes the cosine distance between the masked item's embedding and the fixed reference embedding, ranks the masked versions, and constructs the attribution map by summing the masks from the top and bottom distance percentiles. In 'mirror' mode the bottom-percentile masks are subtracted rather than added. The paper reports that on ImageNet classification vectors and CLIP image-caption embeddings the resulting maps highlight plausible features, that the maps are robust under small input perturbations, and that they change appropriately when model weights are randomized. The central claim is that this distance-ranked mask selection, with or without mirror subtraction, is what lets a saliency method move from single-input classification to pairwise distance explanations without reweighting masks by class activations.

Load-bearing premise

The load-bearing bet is that the random masks whose distance scores are the lowest and the highest are the ones whose uncovered pixels carry the features that determine the distance, and that the artifacts of replacing pixels with a constant color do not mislead the result.

Editorial extensions

If this is right

  • If the method works, any model with an embedding, including black-box models, can be locally explained for pairwise distances as long as a masking function exists for the input modality.
  • The mirror-mode maps provide signed attributions, so they express both what pulls an item toward a reference and what pushes it away, which one-sided saliency maps cannot do.
  • Increasing the number of masks reduces the variance between attribution maps across random seeds, so users can trade compute for stability.
  • Selection thresholds above roughly ten percent of the extreme masks add little information, meaning the default configuration sits near a plateau.
  • Quantitative robustness stays high, with average sensitivity around 0.04 to 0.06, and attribution structure collapses immediately when model layers are randomized, as expected if the explanations track the learned model rather than an image prior.

Reading between the lines

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

  • Beyond the paper, the distance-ranked selection scheme suggests a general recipe for contrastive explanations: any scalar relation between two embeddings, such as Euclidean distance or a learned similarity score, could replace cosine distance and the extreme-rank mask filtering would remain unchanged.
  • Because the paper fixes a baseline imputation that replaces masked pixels with a constant and acknowledges this pushes inputs out of distribution, a natural testable extension is to substitute learned infilling or blur; if out-of-distribution artifacts are substantial, attribution maps should shift systematically under that change.
  • The authors' own parameter sweeps hint that the attribution map is not unique: at low mask coverage the car-versus-bicycle map highlights wheels, while at higher coverage it highlights the car body, which implies the explanation depends on the perturbation distribution and users should report it alongside the map.
  • One could extend the method to a single-input setting by comparing an item to a reference class prototype or anchor, producing explanations of why an item is representative of a class without requiring a second real input.
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

6 major / 5 minor

Summary. Distance Explainer is a post-hoc, local XAI method that explains the cosine distance between two points in an arbitrary embedding space. For a to-be-explained item and a reference embedding, the method generates N random binary masks, embeds each masked input, ranks the resulting masked embeddings by their cosine distance to the reference, selects the masks whose distances fall in the top and/or bottom x% (or their mirror difference), and sums the selected masks to obtain a signed attribution map. The paper evaluates the method on ImageNet classifier vectors (e.g., bee vs. fly, car vs. bicycle) and CLIP image-caption embeddings, using incremental deletion for faithfulness, Average Sensitivity for robustness, and the Model Parameter Randomization Test for model dependence, supplemented by qualitative inspection and a hyperparameter exploration. The authors argue that the method is modality-agnostic and fills a gap left by RISE, S-RISE, and CorrRISE.

Significance. The proposed task, explaining pairwise distances in arbitrary embedding spaces, is timely and relevant, and the paper is honest about limitations, ships an implementation and datasets, and uses standard XAI evaluation frameworks such as Quantus. If the extreme-distance-ranking heuristic were validated against known ground truth or alternative attribution methods, the method would be a useful general extension of RISE. However, the reported evidence does not currently establish the central claim: the selection heuristic is never tested against independent ground truth, the faithfulness result rests on a single un-replicated pair, robustness is reported as two point estimates without uncertainty, and one MPRT mode contradicts the paper's own summary of the results.

major comments (6)
  1. [§2.1 (steps 3–5), §5.3] The core assumption that masks whose masked embeddings fall in the extreme low or high distance ranks contain the features driving the distance is asserted but never validated against ground truth, human annotations, or an alternative pairwise attribution method. All quantitative evaluations reuse the same cosine distance both to construct the attribution map and to measure its effect, so they cannot separate genuine feature importance from artifacts caused by fixed-baseline masking, an out-of-distribution concern the authors acknowledge in §6. I request a direct validation experiment, for example on synthetic images with known feature contributions, on real images with semantic part annotations, or by comparing with S-RISE, CorrRISE, and a suitable random-mask baseline.
  2. [§4.1] The faithfulness experiment is conducted on a single bee-versus-fly pair and reports what appears to be a single deletion trajectory without error bars or multiple random seeds, even though the mask generation is stochastic. Because the deletion order and the evaluation metric both derive from the same distance function, the reported asymmetry between LoDF and HiDF is not sufficient to establish that the highlighted regions are causally important; it could reflect distance artifacts from out-of-distribution masked inputs. Please report aggregated deletion curves over multiple pairs and seeds, with confidence intervals, and compare against a random-mask baseline.
  3. [§4.2] Robustness is reported as two point estimates, 0.06 for bee vs. fly and 0.04 for bee vs. bee, with no confidence intervals, no repeated trials, and no comparison to other explainers or to a null model. This is too thin to support the claim of high robustness and consistency. At minimum, report the distribution of Average Sensitivity across multiple mask draws and data pairs, and state the number of repetitions explicitly.
  4. [§4.3, Table 1] The text states that all modes also show low correlations in Table 1 after first-layer randomization, but the independent-shuffle mode reports correlations of 0.92 at layer 4, 0.67 at layer 7, and 0.61 at layer 8, among others. These values contradict the summary and raise the question of why the explanation remains highly correlated with the original map after randomizing intermediate layers; this needs an explanation or a correction, and it weakens the model-dependence claim for that mode.
  5. [§5.3, Figure 10] Figure 10 shows that distance-decreasing and distance-increasing selected masks produce nearly indistinguishable attribution patterns, which the authors interpret as both measuring similar signals. If true, this undermines the signed interpretation of the mirror mode as separating similarity-contributing from dissimilarity-contributing features; please clarify what distinguishes the two signs in the final map, or temper the claim that the method separates these contributions.
  6. [§5.2–5.3, Table 2] The default hyperparameters (mask count, coverage, resolution, selection threshold, and mirror mode) were selected by visually inspecting the same bee-versus-fly and car-versus-bicycle pairs that are later used in the quantitative evaluation, so the reported results do not control for selection bias. An independent validation set or a sensitivity analysis with held-out pairs would make the quantitative claims more credible.
minor comments (5)
  1. [§2.2] The phrase 'we attempted using ad (with a≈20)' appears to contain a typographical error; presumably 'a·d' or 'a×d' is intended.
  2. [§4.2] The robustness setup is under-specified: it is unclear what perturbation is applied when nr_samples=20 and perturb_std=0.1×255 are used together with deterministic RISE mask generation; please state the perturbation protocol explicitly.
  3. [§4.1, Figures 1 and 2] The captions of Figures 1 and 2 do not define the meaning of 'distance: x%' or the vertical-axis units; please state whether the reported values are percentage changes in Δd or absolute values.
  4. [§4.3] The paper cites [17] as an improved MPRT but still uses the original MPRT implementation; a sentence explaining why the improved version was not used would be helpful.
  5. [§6] The out-of-distribution limitation is acknowledged, but the discussion could state whether any of the reported results, such as the faithfulness curves, are expected to be affected differentially by OOD artifacts; this would help readers interpret the numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Distance Explainer's attribution map is an empirical sum of selected random masks, the faithfulness evaluation is a standard self-consistency check, and the self-citations are implementation-level rather than load-bearing.

full rationale

The derivation chain in this paper is empirical, not deductive. Distance Explainer constructs an attribution map by generating random masks, computing cosine distances of the masked embeddings to a fixed reference embedding, ranking the masks by those distances, and summing the extreme-ranked masks (Sections 2.1 and 2.3). No parameter is fitted to a subset of data and then reported as a prediction; the map is a direct statistical summary of the mask-distance pairs. The faithfulness metric in Section 4.1 removes pixels in the order given by the attribution map and measures the change in the same cosine distance used to generate the map; this is the standard self-consistency definition of faithfulness for saliency methods, and while it cannot independently establish causal or perceptual validity of the attributions, it is not a circular reduction because the deletion curve is not the same computation as the mask ranking. The hyperparameter defaults in Section 5.2 were selected by inspecting the same data pairs used in the reported results, which is a data-snooping concern, but the paper makes no predictive claim that those settings are derived from first principles; overfitting is not circularity. The self-citations (DIANNA [28] and the masking blog [23]) support implementation reuse and the possibility of extending masking to text, tables, and time series, but they do not carry the central claim that Distance Explainer identifies similarity- or dissimilarity-driving features. The extreme-distance mask-selection heuristic in Sections 2.1 and 5.3 is an asserted assumption that is tested through deletion curves and parameter sweeps rather than being defined into truth. The Section 6 caveat that fixed-baseline masking can push inputs out of distribution is a genuine validity threat to the distance measurements, but it is an acknowledged limitation of the perturbation strategy, not a circular step. Accordingly, the paper receives a circularity score of 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim depends on five hand-chosen hyperparameters and six background assumptions. No new physical or conceptual entities are introduced; the mirror mode and rank filtering are algorithmic components, not entities. The most fragile axiom is that extreme-distance masks contain the salient features, which is the basis of the method and is not independently validated.

free parameters (5)
  • mask_count = 1000 (default)
    Number of random masks; chosen as a trade-off between stability and computational cost; stability improves with more masks as shown in Table 3.
  • mask_coverage = 0.5 (default)
    Fraction of pixels kept unmasked in each random mask; the authors find performance varies by data item and recommend sweeping this parameter (Section 5.2).
  • mask_feature_resolution = 8x8 (default)
    Grid resolution of masking superpixels; values 8-16 give the best qualitative results in the paper, while finer resolutions require more masks.
  • selection_threshold = 10% per side (default)
    Fraction of extreme-distance masks retained in the mirror mode; chosen after visual inspection showed little change beyond 10% (Section 5.3).
  • selection_mode = mirror (two-sided)
    Choice of one-sided versus two-sided subtractive selection; the authors chose mirror mode because it uses twice as many masks and reduces noise.
assumptions (6)
  • domain assumption Proximity in the embedding space represents semantic similarity.
    Invoked in Section 2 to define the distance being explained; true for CLIP-style contrastive models but not guaranteed for arbitrary embedding spaces.
  • domain assumption Cosine distance is the appropriate distance metric for the embedded spaces considered.
    Section 2.2 justifies cosine distance over Euclidean distance; the authors acknowledge other metrics may suit CLIP better but do not explore them.
  • domain assumption Random masking with a fixed baseline is a valid perturbation for estimating feature importance.
    Inherited from RISE in Section 2.1; the authors note in Section 6 that fixed-baseline masking can create out-of-distribution inputs.
  • ad hoc to paper Masks whose resulting embedding is extremely close to or far from the reference contain the salient features.
    The core heuristic of distance-ranked mask filtering in Sections 2.1 and 2.3; it is asserted and validated only through qualitative inspection, not against ground truth or baselines.
  • domain assumption For ImageNet models, softmax class-score vectors can be treated as embeddings.
    Section 3.1 acknowledges these are 'not typically considered embedded space vectors' but uses them as a controlled testbed.
  • domain assumption The XAI metrics used (faithfulness, average sensitivity, MPRT) are valid for evaluating a pairwise distance explanation.
    Section 4 relies on Quantus implementations without a task-specific validation that these metrics behave correctly for pairwise attribution maps.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable embeddings with Distance Explainer." pith.science (2026). https://pith.science/paper/KOVJDIID

@misc{pith2026250515516,
  author       = {Pith},
  title        = {Pith review of: Explainable embeddings with Distance Explainer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KOVJDIID}},
  note         = {Machine review of arXiv:2505.15516}
}
read the original abstract

While eXplainable AI (XAI) has advanced significantly, few methods address interpretability in embedded vector spaces where dimensions represent complex abstractions. We introduce Distance Explainer, a novel method for generating local, post-hoc explanations of embedded spaces in machine learning models. Our approach adapts saliency-based techniques from RISE to explain the distance between two embedded data points by assigning attribution values through selective masking and distance-ranked mask filtering. We evaluate Distance Explainer on cross-modal embeddings (image-image and image-caption pairs) using established XAI metrics including Faithfulness, Sensitivity/Robustness, and Randomization. Experiments with ImageNet and CLIP models demonstrate that our method effectively identifies features contributing to similarity or dissimilarity between embedded data points while maintaining high robustness and consistency. We also explore how parameter tuning, particularly mask quantity and selection strategy, affects explanation quality. This work addresses a critical gap in XAI research and enhances transparency and trustworthiness in deep learning applications utilizing embedded spaces.

Figures

Figures reproduced from arXiv: 2505.15516 by the authors.

Figure 1
Figure 1. Incremental deletion on the bee image whose distance to [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Distance change (∆d = dcos(deleted, r) − dcos(e, r)) under incremental deletion on bee vs. fly. Vertical: ∆d (negative means the deletion reduced distance to the reference). Horizontal: deleted pixel percentage [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. MPRT (top-down): First image shows the unperturbed attribution map. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: MPRT (bottom-up): Like Figure 3, but perturbation starts with the first [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Attribution maps using default parameters on image-versus-image pairs. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Attribution maps using default parameters on image-versus-caption pairs. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Attribution map convergence with increasing number of masks (rows) [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Image of a car versus caption “a bicycle”. From left to right, [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: The effect of using different mask feature resolution is shown. The number [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: One-sided mask selection on bee vs. fly. Top row: selecting only distance [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Two-sided “mirror” selection on bee vs. fly. Left to right: increasing per [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 19 canonical work pages

  1. [1]

    Nature Machine Intelligence 5(9), 1006–1019 (9 2023)

    Achtibat,R.,Dreyer,M.,Eisenbraun,I.,Bosse,S.,Wiegand,T.,Samek,W., Lapuschkin, S.: From attribution maps to human-understandable explana- tions through concept relevance propagation. Nature Machine Intelligence 5(9), 1006–1019 (9 2023). https://doi.org/10.1038/s42256-023-00711-8, https://doi.org/10.1038/s42256-023-00711-8

    doi:10.1038/s42256-023-00711-8 2023

  2. [2]

    In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R

    Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity checks for saliency maps. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Ad- vances in Neural Information Processing Systems. vol. 31. Curran As- sociates, Inc. (2018), https://proceedings.neurips.cc/paper_files/paper/ 2018/file/...

    work page 2018

  3. [3]

    Ali, S., Abuhmed, T., El-Sappagh, S., Muhammad, K., Alonso-Moral, J.M., Confalonieri, R., Guidotti, R., Del Ser, J., Díaz-Rodríguez, N., Herrera, F.: Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence. Inf. Fusion99(C) (Nov 2023). https://doi.org/10.1016/j.inffus.2023.101805, https://doi....

    arXiv 2023

  4. [4]

    In: 29th ACM International Conference on Architectural Sup- port for Programming Languages and Operating Systems, Volume 2 (AS- PLOS ’24)

    Ansel, J., Yang, E., He, H., Gimelshein, N., Jain, A., Voznesensky, M., Bao, B.,Bell,P.,Berard,D.,Burovski,E.,Chauhan,G.,Chourdia,A.,Constable, W., Desmaison, A., DeVito, Z., Ellison, E., Feng, W., Gong, J., Gschwind, M., Hirsh, B., Huang, S., Kalambarkar, K., Kirsch, L., Lazos, M., Lezcano, M., Liang, Y., Liang, J., Lu, Y., Luk, C., Maher, B., Pan, Y., P...

    work page arXiv 2024

  5. [5]

    Com- puter49(05), 54–63 (May 2016)

    Bal, H., Epema, D., de Laat, C., van Nieuwpoort, R., Romein, J., Sein- stra, F., Snoek, C., Wijshoff, H.: A Medium-Scale Distributed System for Computer Science Research: Infrastructure for the Long Term . Com- puter49(05), 54–63 (May 2016). https://doi.org/10.1109/MC.2016.127, https://doi.ieeecomputersociety.org/10.1109/MC.2016.127

    doi:10.1109/mc.2016.127 2016

  6. [6]

    Pro- ceedings of the National Academy of Sciences117(48), 30071–30078 (2020)

    Bau, D., Zhu, J.Y., Strobelt, H., Lapedriza, A., Zhou, B., Torralba, A.: Understanding the role of individual units in a deep neural network. Pro- ceedings of the National Academy of Sciences117(48), 30071–30078 (2020). https://doi.org/10.1073/pnas.1907375117, https://www.pnas.org/doi/abs/ 10.1073/pnas.1907375117

    doi:10.1073/pnas.1907375117 2020

  7. [7]

    In: Proceedings of the Twenty-Ninth Interna- 18 Meijer and Bos tional Joint Conference on Artificial Intelligence

    Bhatt, U., Weller, A., Moura, J.M.F.: Evaluating and aggregating feature- based model explanations. In: Proceedings of the Twenty-Ninth Interna- 18 Meijer and Bos tional Joint Conference on Artificial Intelligence. IJCAI’20 (2021)

    work page 2021

  8. [8]

    Cognitive Computation17(1), 19 (12 2024)

    Boselli, R., D’Amico, S., Nobani, N.: explainable ai for word embeddings: A survey. Cognitive Computation17(1), 19 (12 2024). https://doi.org/10. 1007/s12559-024-10373-2, https://doi.org/10.1007/s12559-024-10373-2

    work page doi:10.1007/s12559-024-10373-2 2024

Show all 37 references
  1. [9]

    https://keras.io (2015)

    Chollet, F., et al.: Keras. https://keras.io (2015)

    2015

  2. [10]

    Journal of Ar- tificial Intelligence Research73, 673–707 (Feb 2022)

    Chrupala, G.: Visually grounded models of spoken language - a sur- vey of datasets, architectures and evaluation techniques. Journal of Ar- tificial Intelligence Research73, 673–707 (Feb 2022). https://doi.org/10. 1613/jair.1.12967, dBLP’s bibliographic metadata records provid...

    2022

  3. [11]

    Natural Language Engineering23(1), 155–162 (2017)

    Church, K.W.: Word2vec. Natural Language Engineering23(1), 155–162 (2017). https://doi.org/10.1017/S1351324916000334

    2017 doi

  4. [12]

    In: International Conference on Learning Representa- tions (2021), https://openreview.net/forum?id=YicbFdNTTy

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Un- terthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on...

    2021

  5. [13]

    In: Proceedings of the 25th ACM SIGKDD In- ternational Conference on Knowledge Discovery & Data Mining

    Gade, K., Geyik, S.C., Kenthapadi, K., Mithal, V., Taly, A.: Explain- able ai in industry. In: Proceedings of the 25th ACM SIGKDD In- ternational Conference on Knowledge Discovery & Data Mining. p. 3203–3204.KDD’19,AssociationforComputingMachinery,NewYork,NY, USA (2019). https...

    2019

  6. [14]

    International Journal of Applied Earth Observation and Geoinformation112, 102869 (2022)

    Gevaert, C.M.: Explainable ai for earth observation: A review includ- ing societal and regulatory perspectives. International Journal of Applied Earth Observation and Geoinformation112, 102869 (2022). https://doi. org/https://doi.org/10.1016/j.jag.2022.102869, https://www.scie...

    2022

  7. [15]

    Nature585(7825), 357–362 (Sep 2020)

    Harris, C.R., Millman, K.J., van der Walt, S.J., Gommers, R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor, J., Berg, S., Smith, N.J., Kern, R., Picus, M., Hoyer, S., van Kerkwijk, M.H., Brett, M., Haldane, A., del Río, J.F., Wiebe, M., Peterson, P., Gérard-Marchant, P., S...

    2020 doi

  8. [16]

    Journal of Machine Learning Research24(34), 1–11 (2023), http://jmlr.org/papers/ v24/22-0142.html Explainable embeddings with Distance Explainer 19

    Hedström,A.,Weber,L.,Krakowczyk,D.,Bareeva,D.,Motzkus,F.,Samek, W., Lapuschkin, S., Höhne, M.M.C.: Quantus: An explainable ai toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research24(34), 1–11 (2023), http://jmlr.org/...

    2023

  9. [17]

    Hedström, A., Weber, L., Lapuschkin, S., Höhne, M.M.: Sanity checks re- visited: An exploration to repair the model parameter randomisation test (2024), https://arxiv.org/abs/2401.06465

    2024 arXiv

  10. [18]

    PLOS Computational Biology 17(2), 1–18 (02 2021)

    Huber, F., Ridder, L., Verhoeven, S., Spaaks, J.H., Diblen, F., Rogers, S., van der Hooft, J.J.J.: Spec2vec: Improved mass spectral similarity scoring through learning of structural relationships. PLOS Computational Biology 17(2), 1–18 (02 2021). https://doi.org/10.1371/journa...

    2021 doi

  11. [19]

    Computing in Science & Engineering9(3), 90–95 (2007)

    Hunter, J.D.: Matplotlib: A 2d graphics environment. Computing in Science & Engineering9(3), 90–95 (2007). https://doi.org/10.1109/MCSE.2007.55

    2007 doi

  12. [20]

    Foundations and Trends®in Machine Learning12(4), 307–392 (2019)

    Kingma, D.P., Welling, M.: An introduction to variational autoen- coders. Foundations and Trends®in Machine Learning12(4), 307–392 (2019). https://doi.org/10.1561/2200000056, http://dx.doi.org/10.1561/ 2200000056

    2019 doi

  13. [21]

    In: Tescher, A.G., Ebrahimi, T

    Lu, Y., Ebrahimi, T.: Explanation of face recognition via saliency maps. In: Tescher, A.G., Ebrahimi, T. (eds.) Applications of Digital Image Pro- cessing XLVI. vol. 12674, p. 126740U. International Society for Optics and Photonics, SPIE (2023). https://doi.org/10.1117/12.2677...

    2023 doi

  14. [22]

    In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Lu, Y., Xu, Z., Ebrahimi, T.: Towards visual saliency explanations of face verification. In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4714–4723 (2024). https://doi.org/10.1109/ WACV57701.2024.00466

    2024

  15. [23]

    https://blog

    Meijer, C.: Masking time-series for explainable ai. https://blog. esciencecenter.nl/masking-time-series-for-explainable-ai-90247ac252b4 (2024), accessed: 18 October 2024

    2024

  16. [24]

    ACM Comput

    Nauta, M., Trienes, J., Pathak, S., Nguyen, E., Peters, M., Schmitt, Y., Schlötterer, J., van Keulen, M., Seifert, C.: From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating ex- plainable ai. ACM Comput. Surv.55(13s) (7 2023). https://doi...

    2023 doi

  17. [25]

    Journal of Machine Learning Research 12, 2825–2830 (2011)

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit- learn: Machine learning in Python. Journal of Mac...

    2011

  18. [26]

    In: Proceedings of the British Machine Vision Conference (BMVC) (2018), http://bmvc2018.org/contents/papers/ 1064.pdf

    Petsiuk, V., Das, A., Saenko, K.: Rise: Randomized input sampling for explanation of black-box models. In: Proceedings of the British Machine Vision Conference (BMVC) (2018), http://bmvc2018.org/contents/papers/ 1064.pdf

    2018

  19. [27]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th Int...

    2021

  20. [28]

    https://doi.org/10.5281/zenodo.14337052, https://doi.org/10.5281/zenodo.14337052

    Ranguelova, E., Bos, P., Liu, Y., Meijer, C., Alidoost, F.S., Oostrum, L., Crocioni, G., Jansen, A., Ootes, L., Chandramouli, P., Smeets, S., van der Spek, W.: dianna (Oct 2024). https://doi.org/10.5281/zenodo.14337052, https://doi.org/10.5281/zenodo.14337052

    2024 doi

  21. [29]

    why should I trust you?

    Ribeiro, M., Singh, S., Guestrin, C.: “why should I trust you?”: Explaining the predictions of any classifier. In: DeNero, J., Finlayson, M., Reddy, S. (eds.) Proceedings of the 2016 Conference of the North American Chap- ter of the Association for Computational Linguistics: D...

    2016 doi

  22. [30]

    Nature Computational Science4(1), 43–56 (1 2024)

    Savcisens, G., Eliassi-Rad, T., Hansen, L.K., Mortensen, L.H., Lille- holt, L., Rogers, A., Zettler, I., Lehmann, S.: Using sequences of life- events to predict human lives. Nature Computational Science4(1), 43–56 (1 2024). https://doi.org/10.1038/s43588-023-00573-5, https://d...

    2024 doi

  23. [31]

    In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embed- ding for face recognition and clustering. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 815–823. IEEE (Jun 2015). https://doi.org/10.1109/cvpr.2015.7298682, http://dx.doi.or...

    2015

  24. [32]

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam:Visualexplanationsfromdeepnetworksviagradient-basedlocal- ization. Int. J. Comput. Vision128(2), 336–359 (2 2020). https://doi.org/ 10.1007/s11263-019-01228-7, https://doi.org/10.1007/s11263-...

    2020 doi

  25. [33]

    Shahroudnejad, A.: A survey on understanding, visualizations, and expla- nation of deep neural networks (2021), https://arxiv.org/abs/2102.01792

    2021 arXiv

  26. [34]

    Master’s thesis, Univer- sity of Amsterdam (UvA), Amsterdam, Netherlands (9 2023), available at https://staff.fnwi.uva.nl/a.s.z.belloum/MSctheses/MScthesis_Willem_ van_der_Spec.pdf

    van der Spek, W.: Explaining the Explainer. Master’s thesis, Univer- sity of Amsterdam (UvA), Amsterdam, Netherlands (9 2023), available at https://staff.fnwi.uva.nl/a.s.z.belloum/MSctheses/MScthesis_Willem_ van_der_Spec.pdf

    2023

  27. [35]

    Machine Learning and Knowledge Extraction3(3), 615–661 (2021)

    Vilone, G., Longo, L.: Classification of explainable artificial intelligence methods through their output formats. Machine Learning and Knowledge Extraction3(3), 615–661 (2021). https://doi.org/10.3390/make3030032, https://www.mdpi.com/2504-4990/3/3/32

    2021 doi

  28. [36]

    Information Fusion77, 29–52 (2022)

    Yang, G., Ye, Q., Xia, J.: Unbox the black-box for the medical ex- plainable ai via multi-modal and multi-centre data fusion: A mini- review, two showcases and beyond. Information Fusion77, 29–52 (2022). https://doi.org/https://doi.org/10.1016/j.inffus.2021.07.016, https://www...

    2022 doi

  29. [37]

    In: Proceedings of the 33rd In- ternational Conference on Neural Information Processing Systems

    Yeh, C.K., Hsieh, C.Y., Suggala, A.S., Inouye, D.I., Ravikumar, P.: On the (in)fidelity and sensitivity of explanations. In: Proceedings of the 33rd In- ternational Conference on Neural Information Processing Systems. Curran Associates Inc., Red Hook, NY, USA (2019)

    2019

Pith tools

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