acceptodds
Under review as a conference paper at ICLR 2027

Attribution Upsampling should Redistribute, Not Interpolate

Abstract

Attribution methods in explainable AI rely on upsampling techniques that were designed for natural images, not saliency maps. Standard bilinear and bicubic interpolation systematically corrupts attribution signals through aliasing, ringing, and boundary bleeding, producing spurious high-importance regions that misrepresent model reasoning. We identify that the core issue is treating attribution upsampling as an interpolation problem that operates in isolation from the model's reasoning, rather than a mass redistribution problem where model-derived semantic boundaries must govern how importance flows. We present Universal Semantic-Aware Upsampling (USU), a principled method that reformulates upsampling through ratio-form mass redistribution operators, provably preserving attribution mass and relative importance ordering. Extending the axiomatic tradition of feature attribution to upsampling, we formalize four desiderata for faithful upsampling and prove that interpolation structurally violates three of them. USU emerges as the unique operator satisfying all four. Controlled experiments on models with known attribution priors verify USU's formal guarantees; evaluation across ImageNet, CIFAR-10, and CUB-200 confirms consistent faithfulness improvements and qualitatively superior, semantically coherent explanations.

Primary area:
interpretability and explainable AI
Keywords:
explainable AI, feature attribution, saliency maps, attribution upsampling, axiomatic explainability, explanation faithfulness, mass conservation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tUbSOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1(−2)
Soundness
1(−2)
Read review
Paper summary

The works spots that upsampling is unfaithful for attributions. 4 attributes of attributions are defined (the ones that are upsampled) and 3 of them are not met by interpolation. So the USU is introduced to mitigate this. Perturbation-based evaluation is performed to test the hypothezis and methods using both synthetic and real datasets on CNNs and Vits.

Critical strengths

Upsampling is taken for granted but a bit forgotten for XAI. Good problem to tackle.

Nice intuition and formalization (4 requirements)

Method agnostic

Theoretical analysis which seems to be honest.

Critical weaknesses

USU uses perturbations to identify segments, so wouldn't that mean that USU is a perturbation XAI method mixed with attributions?

No comparison to LayerCAM, CAMERAS, Jitter-CAM, Augmented Grad-CAM and XRAI, even though they are listed. Moreover, there is no comparison to literature baselines such as: AnyUp, FeatUp, JAFAR, LoftUp.

Results: infidelity changes by 4 orders of magnitude, and it is not scale-invariant. So maybe this should be related to normalization and scale issues and not the method?

Table 9 and Table 1 results do not match? should they not match or is the some kind of error as they look similar?

D1 says that each 7×7 GradCAM cell has its own 32×32 pixel block. But it is not true, as this is related to receptive field of the model.

USU-IWMR violates D1

hierarchical pipeline violates D2 and D4

Such statements are unnecessary and annoying: "complete remedy". There is not much chance that you are solving a field. Especially that many other methods exist there.

Questions for authors

How much gain is from perturbations?

How USU compares to other methods such as AnyUp?

Why haven't you compared to methods you described in Appendix?

What time and computation is needed when compared to others?

Why infidelity differs so much between tables?

Why GradCAM is limited to 32x32 as it has whole receptive field?

As some of your variants are violating your own assumptions/requirements, what kind of guarantess do they have? I cannot see them.

Can you show localization on ImageNet?

5 comments
AuthorsOpenReview ↗

Response to Reviewer tUbS (1/2)

As authors, we find it particularly hard to explain our work and our choices in reply to a review that rests on a fundamental misunderstanding of the paper and of the concepts behind it.

Everything seems fundamentally to stem from not understanding, although it is in the title of our paper, that we investigate the upsampling stage in XAI (and the errors that stem from it). As a result, the reviewer misses entirely that:

  1. Intuitively, an upsampling operator (e.g., the one we propose) does not produce any explanation, unlike the XAI methods listed by the reviewer. In turn, this means that an upsampling operator takes an existing explanation; it does not create one. The XAI methods listed by the reviewer are therefore its inputs, not its alternatives.
  2. The methods the reviewer calls "literature baselines" (AnyUp, FeatUp, JAFAR, LoftUp) are feature upsampling methods, not attribution upsampling methods: they upsample the features of a backbone. Where their papers report CAM results (FeatUp, JAFAR), the upsampling is applied to the features inside the attribution method, not to a given attribution map. They therefore address a different problem, which is why our paper does not cite them.

We now address the weaknesses (and misconceptions) listed in the review one by one, in the order in which they appear, since they are not numbered.

  • W1. USU is not a "perturbation XAI method mixed with attributions", because it does not generate any explanation. The operator itself does not use perturbations either: it takes the segment scores as an input. They come from the quality estimator that implements the chosen notion of faithfulness (see our response to Reviewer G3X7), which may or may not use perturbations; masked perturbation, shown in Fig. 4, is one such choice.
  • W2. How can we compare to LayerCAM, CAMERAS, etc., if our technique does not produce any explanation? Our benchmark covers seven XAI methods (Table 9), but their explanations are the input of our method, not a baseline to compare against. Similarly, AnyUp, FeatUp, etc. are feature upsampling methods, not attribution upsampling methods (point 2 above).
  • W3. The difference in Infidelity between Table 9 and Table 1 is due to how the metric is computed: independently sampled in one table, cumulative in the other. Both are common operating modes of XAI evaluation metrics.
  • W4. (D1) doesn't prescribe any number and therefore does not say that each 7×7 Grad-CAM cell has its own 32×32 pixel block. It is defined for any division of the pixels into neighbourhoods, one per cell of the attribution map, whatever their size (Sec. 3.1): 7×7 cells with 32×32 blocks are one case, and ViT-B-16 gives 14×14 cells with 16×16 blocks. Additionally, the size of an attribution map depends not only on the architecture but also on the layer at which the model is probed, which is why even "full-resolution" XAI methods require upsampling when they are applied to intermediate layers (Sec. 2.1).
  • W5. USU-IWMR violates (D1) by design (as opposed to USU, which does not). The paper explains the rationale behind it (Sec. 3.5) why this makes sense when the input attribution map is imperfect, and the inherent tension between being faithful to the model and being faithful to the input attribution map when the two conflict.
  • W6. The desiderata are properties of the operator, and at every step of the hierarchical pipeline the operator applied is USU, which satisfies (D1) to (D4) for the scores it is given (Thm. 5). The detailed proof and derivation are in Appendix A.5. If the reviewer sees an error in them, we ask for the specific line or step.
AuthorsOpenReview ↗

Response to Reviewer tUbS (2/2)

Now we move to the questions.

  • Q1. This question is fundamentally ambiguous: it does not say which gain is meant or how it should be measured. It also presupposes that perturbations are part of USU or the upsampling operator. As we replied to Reviewer G3X7, there is no consensus within XAI on how faithfulness should be measured, so by design USU requires a notion of faithfulness (a quality estimator) to be provided. Any perturbation comes from choosing an estimator that perturbs the input when computing its notion of faithfulness, as the one in Fig. 4 does. The perturbation is then inside the estimator that produces the scores and cannot be separated from the rest of its protocol; a different estimator will often give different results (Appendix B.5). What can be measured is the contribution of the scores as a whole: with equal scores, USU reduces to nearest-neighbour upsampling, so the Nearest rows of Tables 8 and 10 show USU without them (ImageNet, VGG16: 7.15×1067.15\times10^6, against 1.27×1051.27\times10^5 with them), but this is a trivial result.
  • Q2. USU and AnyUp cannot be compared directly, because they address different problems (point 2 above). AnyUp upsamples the features of a vision encoder and is evaluated on semantic segmentation, depth and surface-normal estimation; its paper does not consider attribution maps. If it were nevertheless applied to an attribution map, each output value would be a weighted combination of coarse values, with weights that ignore the segment scores so, like interpolation, it could not satisfy (D2) (Thm. 2).
  • Q3. For the reason given in point 1 and in W2: the methods described in the appendix are XAI methods, whose explanations are the input of an upsampling operator, not an alternative to it.
  • Q4. On top of the coarse map, the cost is T=O(D⋅n)+Q⋅Cq+TsegT = O(D⋅n) + Q⋅C_q + T_\text{seg}, with nn output pixels, QQ calls to the quality estimator, CqC_q the cost of one call and TsegT_\text{seg} the cost of the segmentation. The redistribution alone is O(n)O(n), the same order as interpolation.
  • Q5. See W3: the two tables compute Infidelity in two different modes, independently sampled in one and cumulative in the other, so their values are not on the same scale. We will state the mode in both captions.
  • Q6. See W4: nothing in our work limits Grad-CAM to 32×32. The resolution of a Grad-CAM map is set by the layer at which it is computed, and (D1) applies to whatever resolution that is.
  • Q7. They are in the paper, in four places (see also W5 and W6).
    (i) Sec. 3.5, in the main text, lists the guarantees of Soft IWMR.
    (ii) Thm. 6 and Cor. 1 prove them (Appendix A.3): the total mass is conserved, each neighbourhood sums exactly to its reassigned mass, and (D2) with the reassigned masses, (D3) and a semi-local form of (D4) hold.
    (iii) Table 2 shows which desiderata USU and USU-IWMR satisfy.
    (iv) For the hierarchical pipeline, every step applies USU (Thm. 5), and Thm. 8 proves that the total mass is conserved through the whole recursion (Appendix C.7).
  • Q8. The question does not say whether "localization" means a metric or more qualitative examples. In the paper, and generally within the XAI community, "localization" is a quality estimator (metric which assess explanation fitness), measured against ground-truth attributions (Sec. 4.1), and ImageNet has no ground-truth attributions. Measuring it against fixed annotations says nothing about the model's actual reasoning, and is therefore much less trustworthy than the results we provide on RRL-constrained models, where the ground truth is known (Tables 4 and 5).
Reviewer_tUbSOpenReview ↗

Rebuttal comment

Thank you for your rebuttal.

First of all, I would like to point out that it is job of the Authors to convey the work they are doing. If you believe I have a fundamental misunderstanding of the work (and I am knowledgeable of the field), this work would be completely unclear for people in general AI field.

I have also spent more time of this work and noticed that:

The authors say that with equal scores USU reduces to nearest-neighbour upsampling, so the Nearest rows show "USU without scores". Nearest is worse than Bilinear in every setting of Tables 8 and 10. So the operator on its own adds nothing, and every improvement over interpolation comes from the segment scores.

You ask for "the specific line". Appendix C.7, l. 1609–1611: "(D2) and (D4) with respect to segment scores are not guaranteed" for the fused output.

LayerCAM/CAMERAS/Augmented Grad-CAM "are inputs, not alternatives": Appendix B.2 (l. 1303–1307) presents them as competing refinement strategies. So it means that compute budget is different and comparison is not fair!

D1 defines neighbourhoods as disjoint receptive fields (l. 213–214). Real receptive fields overlap, so cross-block "mass leakage" is not unfaithfulness.

Table 1 vs 9. In the paper there is no distinguishes between "cumulative" and "independently sampled" Infidelity modes (Table 9 caption: "raw infidelity values"). Even the unit-free improvement factors disagree (CUB, GradCAM: ~18× vs ~944×). Scale invariance of Infidelity is still unaddressed.

In all 36 interpolation rows of Tables 8/10: ROAD = PixelFlipping − 0.04, PixelFlipping + IROF = 1.03, fixed offsets vs Bilinear in every setting, and Lanczos-3 = Nearest on all five metrics. Please explain, and without sharing per-image results and code, it is not possible to verify this.

I see no reason to increase my score. Reported gains are not connected to proposed operator, the rebuttal contradicts the paper, and the results in Tables 1, 8 and 9 raise concerns about soundness.

Reviewer_1XyMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper argues that conventional interpolation is inappropriate for upsampling attribution maps because it can violate attribution mass conservation, semantic ordering, and locality. It proposes Universal Semantic-Aware Upsampling, which treats upsampling as semantic mass redistribution and derives a ratio-form operator from four desiderata. Experiments on synthetic and real-world datasets report substantial improvements over standard interpolation.

Critical strengths

Treating attribution upsampling itself as an XAI faithfulness problem is an interesting perspective. The authors explicitly distinguish interpolation from attribution-mass redistribution.

Strong empirical evaluation. (But see weaknesses).

The reported improvements are large and consistent.

Critical weaknesses

D3 is a strong modelling choice. Requiring a score change to have the same multiplicative effect at every absolute score level essentially drives the exponential potential. Thus, the "unique" solution is unique conditional on the chosen axioms, rather than demonstrating that USU is uniquely faithful in a broader sense.

The full pipeline is fairly complicated. Hierarchical segmentation, perturbation-based segment scoring, boundary detection, adaptive mixing, temperatures and redistribution make the practical system substantially more elaborate than the simple ratio rule suggested by the core theory.

The strongest empirical variant intentionally violates D1, the neighbourhood completeness axiom used to motivate the rejection of interpolation. The paper gives a sensible explanation for this, but it creates some tension between the theoretical argument and the empirically preferred method.

There is also contemporary work explicitly pursuing semantic-aware CAM explanations, such as SA-CAM: Semantic-aware visual explanations for deep convolutional networks, making the very broad “semantic-aware explanation” idea less novel even if the particular axiomatic redistribution formulation is. It would be important to put this work into context (in SOTA and maybe in the experimental evaluation).

Questions for authors

"But attribution maps are not images." What's an image? What's an attribution map? In a paper that tries to be formal, it would be interesting to know your definition of these two concepts, more than just saying they are different.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1XyM (1/2)

We thank the reviewer for the careful and constructive reading, and for the interesting points it raises, which we discuss below.

W1. This is an extremely exciting and interesting point, and we agree; however, the issue is related to the field of XAI and its current direction as much as it is to our work. The uniqueness is relative to the desiderata, as the Shapley value is unique relative to its axioms. When first introduced, such axioms were (and still are) highly debated; however, they have slowly become the standard: many others could have been used instead, but those seem to be the ones that we have silently settled for. In this sense, we extended the axiomatic tradition of feature attribution to the upsampling domain and, intuitively, we argue that the upsampling stage should follow the same rules and laws as the rest of the pipeline of an XAI method to ensure "faithfulness".

Because there is no consensus on what faithfulness is and, most importantly, on how it should be measured (hence the plethora of existing XAI methods, and of metrics to rank them), we purposely designed our method so that this notion (and its assumptions!) is decoupled from the more general and applicable theoretical part of our work. This allows future works and discoveries to still make our work, and its insights, still relevant.

Being "uniquely faithful in a broader sense" is a much harder and more fundamental issue, related more to the field of XAI as a whole than to our technique. It would in turn require a proper definition of "faithfulness" and a strict way to assess it, on which there is no established consensus (yet!). In this direction, our work discusses and analyzes a much more subtle and narrow aspect of faithfulness, but one that is more relevant and closer to the goal of our work: whether the upsampling should be faithful to the original explanation or to the model's reasoning. How do we operate when the two conflict? Hence the idea of relaxing (D1), and our variant USU-IWMR. This variant performs very well, and often better than the traditional USU, because existing attribution methods make strong assumptions about what "faithfulness" is, and therefore perform well only when measured with a metric (quality estimator) that ranks along that direction.


W2. While our operator is elegant and simple in its formulation, USU in practice requires more moving parts. The redistribution itself has only one simple parameter, the temperature ε\varepsilon; all the increased complexity comes from the depth and hierarchical refinement machinery, which are all engineering efforts to make it work in practice, in the real world, on real models. As we previously replied to Reviewer G3X7, our main contribution, which in our eyes has a much broader impact, comes from our theory. We sat down and investigated, proved, showed and explained the root of a seemingly forgotten, or taken for granted, step of explainability which, to our surprise, is a source of great errors and systematically, silently corrupts explanations without people knowing it! We went to great lengths to properly evaluate and quantify the error, its origins and its effects; hence USU and its variant, including their extremely thought-out, comprehensive and extensive evaluation. While our theory is generally applicable, we did not try to make USU the fastest, nor to give it its simplest form, which perhaps future work can focus on.

AuthorsOpenReview ↗

Response to Reviewer 1XyM (2/2)

W3. This is the continuation of W1, and refers to a problem which has more to do with how XAI is currently conducted than with our work per se. In the paper we provide a detailed explanation of the issue, but here we can expand on some of the points, which will also be relevant for W4. There is a very specific reason why, in practice, USU-IWMR performed better than the original method when relaxing the constraint: the XAI field offers many attribution methods, each often producing distinct explanations. Each method, implicitly or explicitly, has its own assumptions on what "faithfulness" is (the objective of the method) and on how it should be measured. One of the reasons the community developed a set of metrics, or quality estimators, on top is to rank and try to understand such disagreements; however, to make matters worse, such metrics themselves come with strong assumptions about what is measured and how, and themselves disagree. One (of many) reasons why theoretically faithful attribution methods score well with respect to one metric and worse with respect to another sometimes boils down to their objectives and assumptions not matching. This is a fundamental issue with the landscape of XAI, not so much with our work.

However, because we seek to deepen our understanding, in our paper, as you mentioned, we analyze a similar issue, but one that is more relevant to, and addressable by, our method. How do we act, when upsampling, when the original coarse explanation is imperfect and disagrees with a given notion of faithfulness? This can, arguably, be elegantly addressed by relaxing (D1), giving this control back to the practitioner, who can control "how faithful" to be to the original explanation or to the metrics using a simple temperature parameter, like a volume slider! We note here that there is no "correct" or "wrong" explanation: both are correct and work as intended, they just fulfil different goals.


W4. This follows from W3. SA-CAM is a CAM method that uses the semantic correlation among activation maps: it is semantic-aware in how the coarse map is computed, while USU is in how a given coarse map is upsampled, so it is a possible input of USU, not an alternative to it. Building the semantics into the attribution method itself means adopting the assumptions of that method, which may or may not reflect the intended goal; keeping the upsampling separate avoids this. These are, however, practical considerations. Our main objective was to investigate a forgotten problem, and to study its origins, causes and effects, so as to provide new insights into existing techniques that are taken for granted. One last, subtle point, which also holds for quality estimators: there are only so many ways to actually measure what happens inside a model, so methods will inevitably share practical ingredients; where they differ, and where the novelty of our work lies, is in the theory behind them. We thank the reviewer for the reference, and will add SA-CAM to the related work, with a paragraph on semantic-aware CAM methods.


We now conclude with the questions.

Q1. We fully agree with you and, in fact, definitions and semantics are a bit of a sore point in the XAI community. Given the page limit, we tried, to the best of our abilities, to create a compelling and easy-to-follow narrative that, despite the high amount of theoretical content, has useful insights for different levels of readers, so that anyone, no matter their background, can find something interesting or useful, as we are all progressively coming into more contact with AI. That said, we will add formal definitions of an image and of an attribution map to Sec. 3.1.

We would like to take the time to thank Reviewer 1XyM for the interesting points raised, which genuinely sparked some interesting discussion and gave us the opportunity to provide an additional, longer-term and more authorial view and perspective.

The XAI landscape is saturated with metrics, methods and benchmarks, each promising to address some prior issue or limitation, and often providing very little insight, new knowledge or understanding, which is ultimately what we seek. We could cite many numbers, tables and proofs, of which our paper is full, but it is more relevant and interesting to be able to talk about the problems this way.

Reviewer_G3X7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper contributes to the field of axiomatic explainability by addressing the limitations of using upsampling methods for coarse maps explanations, which can distort visualizations and corrupt their faithfulness. It formalizes four desiderata that upsampling should follow: neighborhood completeness, segment monotonicity, constant conditioning strength, and locality. The authors first show that interpolation methods violate three out of four desiderata, and derive USU, a semantic softmax operator, as the unique operator satisfying all four. They also introduce Soft IWMR as a way to relax the neighborhood completeness desiderata, and a hierarchical boundary refinement procedure. USU is tested on a synthetic dataset of shape classification for a grounded evaluation, and on 3 datasets (Imagenet, CUB-200, CIFAR-10) to measure the method’s faithfulness.

Critical strengths
  1. Clear problem statement and formulation

The paper is well written and well structured, it starts from the limitations of interpolation methods for attribution maps, formalizes what are the properties that should be achieved, and derives the method that satisfies them.

  1. Strong theoretical and empirical foundation

The proposed approach is supported both theoretically, showing that the desiderata properties are satisfied, and empirically, though a ground-truth and fidelity evaluations. This gives a strong foundation to the proposed method

Critical weaknesses
  1. Usability

The paper does not discuss the costs of the proposed technique. Given the complexity of the proposed pipeline (Figure 4 in the Appendix), an overview of usability factors such as the temporal cost, the hardware used, and the number of forward passes in the pipeline are fundamental.

  1. Parts of the appendix are a required read

The appendix contains relevant information that is relevant to fully understand the proposed approach. In particular, Figure 4 shows the full USU implementation pipeline, which clarifies the stages of the method. Some parts of it like the H-Map, are never discussed in the main corpus of the paper. Additionally, the text references Figure 8 multiple times even if it is in the appendix.

Questions for authors

1 - Could the authors clarify if whether the proposed method is class-dependent? How does it perform for images where multiple target classes are present?

2 - The paper reports in total only 4 qualitative images for Imagenet. Could you provide more examples to evaluate the generality of the proposed method?

3 - Concerning weakness 1, which are the average temporal cost, hardware used, and dumber of formal passes?

4 - No code was provided. Will the authors make it available in the future?

1 comment

We first thank the reviewer for the careful reading and constructive suggestions; below, we address the weaknesses and questions raised.

W1. To provide a more useful understanding and analysis of the cost of our method, we provide the computational complexity of our technique; we report counts of operations and estimator calls rather than any time measure (e.g., ms) because the actual running time depends on the explained model and the hardware. In general, on top of the coarse map, the cost is

T=O(D⋅n)+Q⋅Cq+TsegT = O(D⋅n) + Q⋅C_q + T_\text{seg}

with nn output pixels, QQ calls to the quality estimator that produces the segment scores, CqC_q the cost of one call, and TsegT_\text{seg} the cost of the segmentation.

StageEstimator callsOther operations
Interpolation (reference)0O(n)O(n)
Redistribution (USU or Soft IWMR)0O(n)O(n)
Refinement, per level0O(n)O(n)
Scoring1 per scored segmentO(1)O(1) per segment
Whole pipelineP0≤Q≤∑dPdP_0 ≤ Q ≤ \sum_d P_dO(D⋅n)O(D⋅n)

So the redistribution operation alone is of the same order as interpolation. The actual cost will inherently depend on which notion of faithfulness (the quality estimator used by the comparator) that is used. Everything was run on 8xH100 (not needed, but that's the hardware we had available); we will add a section in the appendix with the full hardware specifications the benchmarks were run on.


W2. Unfortunately, due to the page limit, only so much could be included in the main text; in the camera-ready version, we will move Figure 4 into the main text.

However, we would also like to provide an explanation on why the paper was organised this way. While USU and its variant produce quite substantial results (and provide concrete evidence on the impact of such errors), we believe our theory has a much broader impact: it provides a useful set of tools to deepen our understanding of a proven source of error in explainability that we establish formally (Thms. 1 to 3) and that systematically corrupts explanations.


Q1. Our proposed technique and its variant, as an upsampling operator, take no class as input; however, the comparator requires a notion of what "good" is, for which a quality estimator (metric) must be plugged in; depending on the quality estimator (metrics of explanation fitness e.g., faithfulness estimate) a class may be provided as input to the metric.

This choice has been left flexible on purpose: while our theory provides foundationally general applicability, and therefore we believe broader impact, in practice within XAI there's still no consensus on how faithfulness should be measured (Appendix B.5), or on what a "good", that is, more faithful, explanation should look like. Therefore, our method requires, by design, a notion of faithfulness (what is "good") to be plugged in.


Q2. We deliberately include many visualizations from our controlled, RRL-constrained experiments, where the ground truth is known, and fewer from real data, where it is not. Judging such results qualitatively can be misleading, as a map tightly focused on the class of interest might not reflect the model's actual reasoning, and vice versa. This happens very often within XAI as such maps might look clean and convincing, but we truly do not have the knowledge to read them correctly, so they are not a reliable indicator of good results. Towards stricter and provable results, the theory and the quantitative results (1,000 images per dataset, three datasets, three architectures; Tables 1, 8 and 10) hold much more weight than visual inspection alone, since they provide measurable evidence. That said, we will add another page of results in the appendix.


Q3. See W1.


Q4. As mentioned above, while our method shows very significant results (further reinforcing the actual effect and impact of upsampling errors), our main contribution was to formalize and address a so far unexamined aspect of explainability, which is a major source of errors. We believe our theory advances the understanding of a little-known and little-explored part of XAI, perhaps taken for granted, which systematically (and silently) corrupts explanations. That said, if there's interest we will publish the source code upon acceptance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.