- 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
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: , against 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 , with output pixels, calls to the quality estimator, the cost of one call and the cost of the segmentation. The redistribution alone is , 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).
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.
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:
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.