acceptodds
Under review as a conference paper at ICLR 2027

Gradient Difference Enables Efficient Second-Order Vision-Language Interaction Explanations

Abstract

Vision-language encoders such as CLIP excel at connecting visual and linguistic representations, yet interpreting their predictions remains challenging. Existing first-order explanation methods identify important image regions or text tokens individually, but fail to capture cross-modal feature interactions necessary for understanding dual-encoder predictions. Recent studies have therefore moved beyond first-order attribution toward second-order interaction explanations. However, the current frameworks face high computational or memory costs as text lengths and candidate interactions grow. To address these limitations, we propose a lightweight gradient-difference framework that derives second-order interaction explanations directly from existing first-order gradient-based attribution methods. By masking a target textual concept or visual region and measuring the resulting change in matching score gradients, our approach isolates phrase-conditioned visual explanations and region-conditioned textual explanations without introducing specialized interaction machinery or expensive sampling. Applied to standard methods (yielding Grad-CAM+, GAME+, and Grad-ECLIP+), our framework achieves performance comparable to state-of-the-art second-order approaches while significantly reducing computational overhead. The quantitative evaluations and human studies further demonstrate that our approach yields both faithful and user-informative vision-language correspondences.

Primary area:
interpretability and explainable AI
Keywords:
second-order visual explanation, gradient-based attribution, xai

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_F83VOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to efficiently generate fine-grained vision-language interaction explanations for dual encoders such as CLIP. The authors propose a gradient-difference framework that masks a target phrase or image region, computes the difference between the matching-score gradients before and after masking, and applies this difference to Grad-CAM, GAME, and Grad-ECLIP, yielding Grad-CAM+, GAME+, and Grad-ECLIP+. The framework is evaluated on ImageNet, MS COCO, and Flickr30k using phrase-targeted deletion/insertion, the Pointing Game, computational cost, and user studies. The results show that Grad-ECLIP+ generally achieves performance comparable to existing second-order methods while substantially reducing computational and memory costs.

Critical strengths
  1. The problem is important and practically relevant. As discussed in the Introduction, ordinary first-order attribution can identify which image regions or text tokens are important to the overall image-text similarity, but it cannot further determine which image region is associated with a particular phrase. Understanding such fine-grained relationships is valuable for assessing whether the model relies on meaningful vision-language correspondences.

  2. The efficiency results are convincing. Table 1 shows that, under the evaluated configurations, Grad-CAM+, GAME+, and Grad-ECLIP+ require substantially less processing time and memory than exCLIP and FIxLIP. Tables 2 and 5 further indicate that this efficiency improvement does not substantially compromise phrase-targeted deletion/insertion or localization performance.

  3. The core idea is clear and easy to understand. The intuition is straightforward: if a textual concept affects the contribution of certain image regions to the matching score, masking that concept should change the gradients of those regions. The proposed framework directly reuses existing gradient-based explanation methods and provides a substantive and memorable insight rather than a complicated recipe composed of many modules.

Critical weaknesses
  1. Gradient difference may reflect attribution redistribution rather than semantic correspondence. For example, after masking "car," the model may reallocate more matching contribution to "dog," producing a large negative gradient difference in the dog region. This indicates a competitive dependency between the two concepts, but it does not imply that the dog region semantically corresponds to "car." The paper also does not explain how M(A,ΔG)M(A,\Delta G) handles negative values: ReLU would discard competitive interactions, whereas taking the absolute value could incorrectly highlight non-target regions. This issue concerns Sections 3.2-3.3 and Equations (5) and (7).

  2. The current experiments do not rule out this alternative explanation. Tables 2 and 5 mainly test whether high-scoring regions localize the target object, but they do not report signed gradient changes in target and non-target regions. Consequently, these results cannot distinguish semantic correspondence from global attribution redistribution. The authors should report signed interaction maps and separately analyze positive and negative gradient differences in target and non-target regions.

Questions for authors
  1. How does M(A,ΔGP)M(A,\Delta G_P)handle negative gradient differences in Grad-CAM+, GAME+, and Grad-ECLIP+? Does the implementation use ReLU, absolute values, clipping, or min-max normalization? Please clarify the exact processing used for Figures 1 and 3 and Tables 2 and 5.

  2. What are the results in Tables 2 and 5 when explanations are generated separately using $ \Delta G_P^+=\max(\Delta G_P,0)$ and $ \Delta G_P^-=\max(-\Delta G_P,0)?$ Do the positive and negative components have different interpretations?

The text and token labels in Figure 3 are too small to read comfortably at the normal page size. I suggest increasing the font and overall visualization size.

2 comments
AuthorsOpenReview ↗

Rebuttal to reviewer F83V (Part 1)

How does handle negative gradient differences in Grad-CAM+, GAME+, and Grad-ECLIP+?

(1) The sign handling in our paper:

We will clarify the sign handling in Eq. (7). After calculating the gradient difference, the ΔGP\Delta G_P substitutes the gradient used in the standard Grad-ECLIP, GradCAM, and GAME, where Grad-CAM+ and Grad-ECLIP+ apply ReLU after channel aggregation, and GAME+ clips negative attention-weighted responses before propagation. Therefore, the final attribution maps are all based on positive gradients, which highlight regions/words that contribute positively to the matching score. In Figures 1 and 3, the heat maps are min-max normalized to 0 to 1 before visualization. In Table 2, the deletion and insertion ranks the patch values in descending order. In Table 5, the peak location and energy inside the target mask are calculated on the positive heatmaps.

The relationship between positive gradient differences and the use of ReLU:

For positive and negative gradient differences, we should clarify that using ReLU does not mean discarding negative gradient differences or competitive interactions, since both positive and negative gradient differences are multiplied by the feature representations, and ReLU is adopted at the outermost calculation, after the channel aggregation, as in the standard Grad-ECLIP and Grad-CAM. The usage of ReLU aims to keep only the matched patch-token interactions that contribute positively to the image-text matching score. Therefore, our method takes both the positive and negative gradient differences into the patch-token interaction calculation. Then, the results in the experiments below show the evidence that, in the gradient differences, the positive component is more closely aligned with visual evidence for the target phrase, while the negative component can suppress the responses of regions that support the remaining caption.

AuthorsOpenReview ↗

Rebuttal to reviewer F83V (Part 2)

What are the results in Tables 2 and 5 when explanations are generated separately using ΔGP+=max(ΔGP,0)\Delta G^+_P = max(\Delta G_P, 0) and ΔGP−=max(−ΔGP,0)\Delta G^-_P = max(-\Delta G_P, 0)? Do the positive and negative components have different interpretations?

We thank the reviewer for raising the distinction between the effects that positive and negative gradient differences may produce on the attribution map.

We will add a new Appendix section to the revised paper to discuss the effects of the positive and negative components in the gradient differences. To investigate the effects of the positive and negative gradient differences, we apply the ΔGP+\Delta G^+_P and ΔGP−\Delta G^-_P to the Grad-ECLIP+ and evaluate the deletion, insertion, and pointing game as suggestions. In the deletion and insertion metrics, both the target phrase and the caption without the phrase are used to calculate the score changes.

Table A. Deletion and insertion AUC for Grad-ECLIP+ on ViT-B/32 using the positive gradient component (GP+G_P^+).

Scoring textFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
Phrase0.38370.89170.33590.89580.06681.0819
Caption without phrase0.54210.62930.77390.62200.55970.4570

Table B. Deletion and insertion AUC for Grad-ECLIP+ on ViT-B/32 using the negative gradient component (GP−G_P^-).

Scoring textFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
Phrase0.88760.38650.89510.34191.09460.0660
Caption without phrase0.64070.55330.61950.78350.50720.6831

Table C. Interaction recognition for Grad-ECLIP+ on ViT-B/32 using each gradient component separately.

Gradient componentFlickr30k PGA ↑Flickr30k PGE ↑MS COCO PGA ↑MS COCO PGE ↑ImageNet PGA ↑ImageNet PGE ↑
Positive (GP+G_P^+)77.57%59.28%86.01%63.91%92.59%69.35%
Negative (GP−G_P^-)27.42%24.97%22.63%19.09%3.30%4.30%

Tables A, B, and C show that the positive and negative gradient differences prioritize different regions. Across all three ViT-B/32 datasets, maps generated from GP+G_P^+ achieve lower deletion AUC and higher insertion AUC for the target phrase, as well as higher pointing-game accuracy and energy within the annotated target region. When scored against the caption without the phrase, maps generated from GP−G_P^- achieve better deletion and insertion AUCs. These results indicate that the positive gradient component is more closely aligned with visual evidence for the target phrase. The negative component can highlight regions that support the remaining caption, consistent with attribution redistribution after masking.

The text and label size in Figure 3 are too small.

Thanks for pointing that out. We will enlarge the font size and visualization when more space is available if preparing the final version.

Reviewer_m3sJOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method to explain similarity scores of vision-language models by providing cross-modal feature interactions.

The method can be used to extend first-order gradient-based explanations to feature interactions by evaluating the difference of first-order attribution at the explained instance and a masked instance, following a "leave-one-out"-methodology. This provides an interaction score between the masked part and the underlying attribution. This can be viewed as applying a discrete derivative on top of a first-order gradient method, which yields a second-order like interaction score structure.

The authors then compare their method on vision-language models with recently proposed vision-language explanations, FIxLIP and exCLIP, and apply their framework on first-order methods GAME, Grad-CAM and GRAD-ECLIP. The main advantage of this method is the computational complexity, since computing first-order attributions at two instances is much faster than computing second-order derivatives (exCLIP) or second-order discrete (perturbation-based) interactions (FIxLIP).

The methods are evaluated via insertion-deletion scores where the scores are computed on the image heatmap using the masked text sequence, which favors the present method. Additionally, the author provide a user study, which evaluates whether the explanations align with human expectation.

Critical strengths
  1. Effective and practical approach: Computing first-order attribution at masked inputs provides a natural way of measuring interactions, and is much cheaper to compute than second-order interaction measures exclusively relying on second-order derivatives (e.g. exCLIP) or second-order discrete derivatives (e.g. FIxLIP). This idea has also shown to be effective when combined with Shapley values on the masked inputs across many applications [1], including vision-language models.

[1] Baniecki, H., Biecek, P., & Fumagalli, F. (2026). Attributions All the Way Down? The Metagame of Interpretability. arXiv preprint arXiv:2605.06295.

Critical weaknesses
  1. Novelty and technical contribution: This method follows the same idea as the metagame [1] which defines a cooperative game νi(S)\nu_i(S) on the SS-masked inputs and first-order attributions for feature ii (the image features in this case). The corresponding meta-attributions have also been applied to vision-language models. Meta-attributions propose the Shapley value on this metagame as the attribution which gives the interaction measure, and the presented method here is a simpler construction, which evaluates νi(full text)−ν(P-masked text)\nu_i(\text{full text}) - \nu(\text{P-masked text}) as a variant of leave-one-out attribution. An earlier work has also provided a theoretical construction called "sum of power" [2]. Evaluating first-order attributions at masked inputs is therefore not a novel idea, and the current main competitors are not compared. Moreover, the presented method does not further investigate the interpretation of this approach. For instance, [1,2] link this construction to existing perturbation-based interaction measures. It would be desirable to understand how this simplified construction compares with the more powerful Shapley value construction in [1]. The presented masking difference here can be viewed itself as a first-order attribution method when applied to the prediction, which is known to be limited in practice (i.e. leave-one-out style attribution). This, and other axioms such as efficiency, has led to more complex solutions, such as Shapley values (as used in meta-attributions [1]), which are, however, harder to compute in practice and have to be approximated. It currently remains unclear, why in this setting the proposed methodology is preferable. Overall, the construction is quite simplistic, and not sufficiently theoretically supported or compared with existing methods.

  2. Evaluation protocol: The current evaluation protocol heavily favors the presented methodology by scoring the attributions via insertion and deletion curves on the input where only PP is present. Other evaluation methods, e.g. of FIxLIP (Baniecki et al., 2025), which introduce insertion and deletion curves based on the most-similar mask of size kk identified through the explanation, or the pointing game evaluation used by both exCLIP and FIxLIP are not considered. Moreover, the meta-attribution counterparts [1] should be compared, as the approach gives a similar second-order extension of gradient-based methods that can be used for vision-language applications.

Minor: Citation-styles are sometimes off, using \citet when \citep would be needed (e.g. line 302, 304)

[2] Lundstrom, D., & Razaviyayn, M. (2023, July). A unifying framework to the analysis of interaction methods using synergy functions. In International Conference on Machine Learning (pp. 23005-23032). PMLR.

Questions for authors
  1. Why is the leave-one-out attribution beneficial over Shapley-style attribution used in [1]? While it computationally more efficient, what are its limitations?
  2. How does your approach compare with meta-attributions [1] empirically?
  3. What is the corresponding concept that is obtained when applying the masking approach also to as the first-order attribution?
  4. How does your method compare with baselines on other evaluation protocols?
5 comments
AuthorsOpenReview ↗

Rebuttal to reviewer m3sJ (Part 1)

Compare our gradient-difference-based explanation with the Shapley-style meta-attributions in [1], and the leave-one-out attribution map:

Thank you for bringing this relevant work, which we had previously missed as it will appear in NeurIPS 2026 next month. To compare the metagame and our gradient-difference method, we use Grad-ECLIP+ and Meta-Grad-ECLIP as an example. Our method applies the Grad-ECLIP operator after subtracting gradients, while the published Meta-Grad-ECLIP takes coalition marginals of the first-order Grad-ECLIP attribution maps. The difference is substantive. Our raw gradient difference is the grand-coalition marginal of the target phrase in a gradient-valued game. It answers a specific question about the observed caption: how does image-feature sensitivity change when this phrase is removed, and the rest of the caption is retained? The activated gradients are assigned to the target phrase, which is different from the leave-one-out on Grad-ECLIP maps, which assigns part of the first-order attributions. Then, as for comparing the leave-one-out strategy with the Shapley-style strategy, if two phrases and a gradient component that activates only when both are present, leave-one-out assigns the full joint effect to each phrase, while Shapley assigns half to each. The practical advantage of leave-one-out is the much lower cost of answering a specific target in the inputs.

To comprehensively compare Grad-ECLIP+ and Meta-Grad-ECLIP and show the performance difference between our method and the simplified leave-one-out construction of Meta-Grad-ECLIP, we conduct experiments under these three variants: (a) Grad-ECLIP+; (2) Leave-one-out on Grad-ECLIP maps; (3) Meta-Grad-ECLIP. As shown in Table A, since the time consumption of Meta-Grad-ECLIP grows exponentially with the number of tokens, we run the following evaluation over the subset of Flickr30k (1989 samples), MS COCO (2000 samples), and ImageNet-1k (2000 samples) with samples that have a text token length shorter than 12 (<= 2^12 combinations when computing Shapley values and interaction indices in Meta-Grad-ECLIP).

Table A. Processing time per image-text pair on ViT-B/32 with 49 image patches and varying numbers of text tokens (N).

NLeave-one-out Grad-ECLIPMeta-Grad-ECLIPGrad-ECLIP+
100.016 s14.8 s0.015 s
200.019 s4.2 h0.02 s

Table B. Comparison of the faithfulness of phrase attribution maps on ViT-B/32, measured by deletion and insertion AUC.

MethodFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
Leave-one-out Grad-ECLIP0.4690.6830.4940.9460.2920.956
Meta-Grad-ECLIP0.4540.7280.4620.9470.2790.937
Grad-ECLIP+0.4510.6930.4800.9390.2650.972

As shown in Table B, Grad-ECLIP+ does not equal the Leave-one-out Grad-ECLIP, and Grad-ECLIP+ obtains better faithfulness results in most metrics across different datasets. Compared with Meta-Grad-ECLIP, Grad-ECLIP+ has comparable performance, with three higher and three lower, while being substantially (980x to 756,000x) faster. Together with the comparison in the following interaction recognition performance evaluated by Pointing Game, the results demonstrate the practical application value of the proposed method.

AuthorsOpenReview ↗

Rebuttal to reviewer m3sJ (Part 2)

The current evaluation protocol heavily favors the presented methodology by scoring the attributions via insertion and deletion curves on the input where only is present.

We argue against this point because the proposed evaluation protocol is measuring the faithfulness of patch-word attribution results, and other 2nd-order methods, exCLIP and FixLIP, also produce the patch-word interaction value matrix the same as ours. The phrase-targeted patch heat map is obtained by aggregating the attribution map of each token in the phrase. We evaluate all the methods under the same setting, which produces a fair comparison for the word-patch faithfulness evaluation.

Comparison with baselines on other evaluation protocols:

We discuss in detail the reasons that we did not adopt insertion/deletion as in exCLIP or FIxLIP in the main paper Line 314-320, as both of them evaluate the importance of patches and tokens to the whole image-text matching rather than the correlation between image regions and parts of the sentence. Although the deletion and insertion metric in FixLIP focuses more on how the patch-token combinations relate to the whole image-text matching, we reproduce the experiment using the patch-token interaction values obtained from exCLIP, FixLIP, Meta-Grad-ECLIP, Grad-ECLIP+, Grad-CAM+, and GAME+ over the same subset as in Table B: Flickr30k (1989 samples), MS COCO (2000 samples), and ImageNet-1k (2000 samples). The masked image and text matching score is normalized and recorded when deleting/inserting a percentage of patches and tokens at each step, with the selected patches and tokens maximizing the interaction values.

Table C. Deletion and insertion AUC on ViT-B/32 under the FIxLIP evaluation protocol.

MethodFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
exCLIP0.120.34-0.321.27-0.140.45
FIxLIP-0.090.63-0.291.56-0.350.77
Meta-Grad-ECLIP-0.080.59-0.571.43-0.180.56
Grad-ECLIP+-0.020.590.281.55-0.100.56
GAME+-0.080.58-0.311.12-0.030.60
Grad-CAM+0.170.260.570.830.010.38

As the results shown in Table C, except for FixLIP, which mostly produces the best performance, the method with second-best performance varies for different metric and dataset. Since the matched patches and tokens with high-interaction values may be deleted/inserted at the same step, there is uncertainty about whether the remaining patches and tokens can also form good matches, which causes large fluctuations in the score. Therefore, we think that this protocol is not suitable for evaluating the patch-token correlation or the region-phrase matching relationship.

As for the pointing game, the evaluation in Table 5 on 5000 samples constructed from ImageNet-1k is the same metric as in exCLIP, FIxLIP, and Metagame under the 4-object setting. Here we present the Point Game results under the other settings (1 object, 2 objects, 3 objects) and 4 objects on a subset of 500 images due to the high time cost of the game theory methods (FIxLIP and Meta-Grad-ECLIP).

Table D: Interaction recognition performance when evaluating Point Game Accuracy and Energy (PGA and PGE) under 1 object, 2 objects, 3 objects, and 4 objects in the caption.

Method1 object PGA ↑1 object PGE ↑2 objects PGA ↑2 objects PGE ↑3 objects PGA ↑3 objects PGE ↑4 objects PGA ↑4 objects PGE ↑
exCLIP0.610.440.770.620.840.700.860.73
FIxLIP0.840.660.840.640.840.640.850.64
Meta-Grad-ECLIP0.890.650.870.660.890.670.880.68
Grad-ECLIP+0.880.620.870.660.860.670.860.67
GAME+0.890.540.880.540.870.540.850.54
Grad-CAM+0.890.680.890.730.890.770.890.75

Table D shows the comparison results with ViT-B/32. The gradient-difference-enabled Grad-CAM+ obtains the best performance under most conditions. The performance of Grad-ECLIP+ is the same as Meta-Grad-ECLIP at 2-object PGA, PGE, and 3-object PGE, while slightly worse under other conditions (1-object: 0.88 vs. 0.89, 0.62 vs. 0.65; 4-object: 0.86 vs. 0.88, 0.67 vs. 0.68). Considering the large time cost and number of token limitations caused by the Shapley value calculation in Meta-Grad-ECLIP, Grad-ECLIP+ offers greater practical value, achieving comparable performance with negligible computational overhead.

AuthorsOpenReview ↗

Ask for a clarification about Question 3.

About the Question 3: What is the corresponding concept that is obtained when applying the masking approach also to as the first-order attribution?

Sorry, I can not make sure of the meaning. Could you please provide more clarification so we can address it accurately?

Reviewer_VVvrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explains CLIP’s image–text matching score by reusing an existing first-order explainer. For a chosen phrase, the method zeros that phrase’s tokens, subtracts the image-feature gradient of the masked caption from the gradient of the full caption, and feeds the difference into Grad-CAM, GAME, or Grad-ECLIP. The paper treats this difference as a cheap stand-in for a cross-modal second derivative. The resulting maps are far cheaper than FIxLIP and exCLIP. Grad-ECLIP+ ranks above the methods that scale to ViT-B/16, while FIxLIP remains stronger on ViT-B/32.

Critical strengths

The method builds a phrase-conditioned map by zeroing the chosen phrase, subtracting the two image-side gradients, and passing the difference through Grad-CAM, GAME, or Grad-ECLIP. Because this map reuses an explainer that already exists, the method needs only two backward passes and does not introduce a separate interaction solver. This design is what keeps the maps computable at ViT-B/16, where FIxLIP is too expensive to run.

Critical weaknesses
  1. The method zeros the phrase tokens and subtracts the two image-feature gradients, and the paper reads that subtraction as a cross-modal second derivative. That reading holds only when masking acts as a small step in the text representation. Zeroing tokens instead produces a large change through a nonlinear text encoder, and the derivative is taken only on the image side. The map therefore remains a conditional first-order gradient and does not provide the second-order interaction the paper claims.

  2. The original explainer collapses this gradient difference into one heatmap for the masked phrase, and the deletion and insertion tests score that heatmap by its similarity to the phrase alone. These tests never measure the change in the full-caption match, and the method never builds a patch–token interaction. On ViT-B/32, where FIxLIP does build those interactions, FIxLIP stays ahead of Grad-ECLIP+. The comparison therefore does not show that the gradient difference recovers the claimed interaction.

Questions for authors

If the same maps are scored by the change in the full-caption match, rather than by similarity to the isolated phrase, does the ranking against exCLIP remain the same? This asks whether the gradient difference explains the prediction the paper starts from.

2 comments
AuthorsOpenReview ↗

Rebuttal to Reviewer VVvr (Part 1)

1. The derivative holds only when masking acts as a small step in the text representation, while zeroing tokens introduce a large change. The derivative is only taken on the image side, so the map therefore remains a conditional first-order gradient and does not provide the second-order interaction.

A: Let u(A) be the normalized image embedding from image features A, and let v(T) be the normalized text embedding. We try to address your concern from the following points:

(1) From the implementation aspect, the image feature and text feature are calculated from two separate encoder branches; starting from the score S(A,T)=u(A)v(T), we cannot obtain the second-order gradients δ2S/δTδA\delta^2 S / \delta T\delta A by backpropagation directly, because the token vectors are not involved in the calculation graph of image features. Therefore, we use the derivation in Sec. 3.2 and Appendix A to approximate the influence of a specific part of the sentence T on the matching score, and further how it relates to the image feature, the δ2S/δTδA∗δT\delta^2 S / \delta T\delta A * \delta T, to the ΔGP\Delta G_P, which can be easily obtained and experimentally proven effective.

(2) The second-order interaction means the patch-token correlation instead of the explanation at the image-caption level, which is not related to the order of the used gradient. That means 2nd-order interaction does not specifically require using a 2nd-order gradient.

(3) A symmetrical derivative can be done on the text side. By removing the visual information from the image patch, a patch-corresponding importance vector can be calculated for the text. We show the bidirectional interaction visualization result in Figure 4 (a).

2-1: The deletion and insertion tests never measure the change in the full-caption match.

A: The purpose of our phrase-targeted deletion/insertion is to evaluate the second-order explanation (patch-word relationship), which measures how well the map identifies patches that support the specific phrase in the caption. If we use the heatmap corresponding to a phrase and the whole caption as text for similarity scores in the test, the evaluation is not accurate enough because wrong patch-phrase interactions may cause score changes due to the highlighted regions being misaligned with some other words in the caption. We then cannot be certain whether the heatmap regions are really aligned with the specific phrase or some other words in the caption.

2-2:On ViT-B/32, where FIxLIP does build those interactions, FIxLIP stays ahead of Grad-ECLIP+. The comparison therefore does not show that the gradient difference recovers the claimed interaction.

A: FIxLIP gives an optimal patch-word alignment result by solving the high computational-cost game-theory problem, while Grad-ECLIP+ offers a much lower-cost way based on existing first-order explanation methods to generate phrase maps and achieves better deletion/insertion results than the other second-order explanation design, exCLIP. In Table 5, Grad-ECLIP+ outperforms FIxLIP on the pointing game metric for both COCO and Flickr30k. Furthermore, in the more important real human user study, the interaction map generated by Grad-ECLIP+ shows higher plausibility and comparable usefulness to FIxLIP (through forward simulation study), demonstrating that gradient-difference-enabled second-order explanations can efficiently and effectively explain interactions between patches and words in image-caption matching. In other words, Grad-ECLIP+ and FIxLIP both improve users' understanding of CLIP, while Grad-ECLIP+ is faster to compute and thus more amenable to real-time explanation that are useful for interactions.

AuthorsOpenReview ↗

Rebuttal to Reviewer VVvr (Part 2)

3. If the same maps are scored by the change in the full-caption match, rather than by similarity to the isolated phrase, does the ranking against exCLIP remain the same?

A: We conduct the experiments as requested: the same phrase-specific maps are scored by the change in the full-caption match, and the results are shown below:

ViT-B/32

MethodFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
FIxLIP0.430.770.430.830.410.66
exCLIP0.600.620.580.700.530.56
Grad-ECLIP+0.480.710.490.760.440.63

ViT-B/16

MethodFlickr30k deletion ↓Flickr30k insertion ↑MS COCO deletion ↓MS COCO insertion ↑ImageNet deletion ↓ImageNet insertion ↑
exCLIP0.570.750.580.790.460.54
Grad-ECLIP+0.530.770.520.820.390.63

As the results shown in the tables, considering the score changes in the full-caption matching, using the phrase maps to perform image deletion and insertion, Grad-ECLIP+ still results in lower deletion and higher insertion AUCs compared with exCLIP. However, for the reasons stated in Answers for 2-1, we still believe that using the full caption to test the patch-token interaction heatmap for a specific phrase in the caption is not appropriate.

Comments on the paper

1 comment
AuthorsOpenReview ↗

An updated version of paper and looking forward to your responses or further suggestions/comments!

Dear reviewers:

Thanks again for all of your constructive suggestions, which have helped us improve the quality and clarity of the paper!

We have uploaded a revised paper and appendix to address the reviewers' insightful comments. The modifications are marked in blue text in the paper and appendix. The major changes are:

  • Added sentences for further clarification about deriving the second-order interaction via the approximation of second-order derivatives in Appendix A (Reviewer VVvr).
  • Added the introduction of the related work, Meta-Grad-ECLIP, in Sec. 1Introduction, and Sec. 2 Related Works (Reviewer m3sJ).
  • Added the computational cost comparison with Meta-Grad-ECLIP in Table 1 (Reviewer m3sJ).
  • Added a comparison with Leave-one-out Grad-ECLIP and Meta-Grad-ECLIP in Appendix D.1 (Reviewer m3sJ).
  • Added a comparison with baselines on the insertion and deletion metric protocol as in FIxLIP in Appendix D.2 (Reviewer m3sJ).
  • Added a comparison with baselines on the multi-object point game as in FIxLIP and Metagame in Appendix D.3 (Reviewer m3sJ).
  • Added a statement about the sign handling in Sec. 3.3 (Reviewer F83V).
  • Added the investigation into the effects of the positive and negative components in the gradient differences in Appendix E (Reviewer F83V).

Please don’t hesitate to let us know if there are any additional clarifications or experiments that we can offer, as we would love to convince you of the merits of the paper. We appreciate your suggestions. Thanks!

-the authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.