acceptodds
Under review as a conference paper at ICLR 2027

Label-Agnostic Attribution for Interpretabilit

Abstract

Attribution maps help users inspect which parts of an image influence a classifier's prediction. They usually explain a chosen class, yet an image can support several plausible labels. In such cases, users may want to inspect the prediction before deciding which class to explain. We develop an attribution method that uses a common objective across all classes: a symmetric measure of how far their predicted probabilities depart from a uniform distribution. Sampled classes guide perturbations of the input, while gradients of this shared objective determine feature importance. This separates the class used to explore an image from the criterion used to explain it. Across the convolutional classifiers and confidence groups studied, restoring highly ranked image regions recovers class scores more effectively than the compared methods. Other methods achieve stronger score suppression when regions are removed. These findings support distribution level attribution as a complement to maps for individual classes, with its benefits depending on how feature importance is evaluated.

Primary area:
interpretability and explainable AI
Keywords:
label agnostic attribution, visual interpretability, image classification, saliency maps, feature importance, distribution level attribution, explainable AI

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_cR9TOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Label-Agnostic Attribution for Interpretability (LAAI). It explains a symmetric objective, KL divergence from a uniform distribution to the classifier’s output probabilities. Sampled classes guide input exploration. Gradients of the shared objective are accumulated using displacements from the original image. Experiments compare attribution rankings through insertion, deletion, and distributional metrics. The main CNN results show stronger insertion performance, with weaker deletion performance than some baselines.

Critical strengths
  1. The explanation target is clear and potentially useful. A shared, class-symmetric objective offers a complementary explanation when users have not selected a class to explain.

  2. The main insertion results are promising. LAAI achieves the highest mean insertion scores across all six CNN–confidence settings in Tables 1–3. The paper also explicitly reports its deletion trade-offs and theoretical limitations.

Critical weaknesses
  1. The core results are unclear. I understand that the exploration applies another softmax when the model already outputs probabilities. And I am confused why each example uses the entire batch target vector rather than its own target. Moreover, appendix C reports a missing evaluation dependency and module error which are unsolved. The paper does not identify a complete runnable configuration for Tables 1–3. These issues leave the main empirical claims insufficiently supported.

  2. The algorithmic contribution is not isolated. The comparisons change both the explanation target and the attribution procedure. The paper lacks direct-gradient and Integrated Gradients baselines for the same KL objective. The added value of the proposed exploration and accumulation is therefore unclear.

Questions for authors
  1. Which exact runnable configuration produced Tables 1–3?

  2. What does the proposed procedure add over direct gradients or Integrated Gradients of the same KL objective?

3 comments

W1 / Q1: Exact configuration behind Tables 1–3.

The experiments evaluate pretrained Inception-v3, ResNet-50, and VGG16 models provided by torchvision on a fixed set of 1,000 images. Inputs have a resolution of 224×224224\times224 and use ImageNet normalization, with all models in evaluation mode. Attribution settings are shared across models: M=20M=20 explorations, T=20T=20 iterations per exploration, perturbation radius ϵ=0.05\epsilon=0.05, and a batch size of 64.

We have provided the complete evaluation entry point and required dependencies; see Appendix C.2.

W2 / Q2: Added value over direct gradients or IG of the same KL objective.

LAAI adds class-guided exploration and displacement-weighted accumulation beyond the choice of a KL objective. Direct gradients use the local derivative at the original image; IG integrates along the straight path from a black reference image to the input. LAAI explores directions guided by sampled classes, evaluates the shared KL gradient at the explored locations, and accumulates attribution weighted by displacement from the original image (Section 3.3). Section 3.4 further analyzes the distinction between this accumulation and path integration.

The added value of exploration and accumulation is demonstrated by comparisons with the same objective and gradient budget. With the same KL objective and 800 gradient evaluations per attribution run, LAAI achieves significantly higher insertion scores than IG800 on all three models, demonstrating more effective ranking of features that support recovery of the original prediction.

The table reports LAAI minus IG800 with adjusted confidence intervals; higher ΔINS and lower ΔDEL are better.

ModelΔINS ↑ΔDEL ↓
ResNet-50+0.1515 [0.1335, 0.1695]+0.0253 [0.0195, 0.0311]
Inception-v3+0.0986 [0.0815, 0.1158]+0.0467 [0.0381, 0.0557]
VGG16+0.1046 [0.0887, 0.1208]+0.0196 [0.0155, 0.0239]

We sincerely thank the reviewer for their valuable feedback. We hope that our responses have addressed your concerns. As the rebuttal period is coming to an end, we would be happy to further discuss any remaining questions or concerns. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your assessment of our paper in light of our clarifications.

Thank you for the rebuttal and the additional IG800 experiments. I appreciate these improvements, but three concerns remain

  1. Why does the exploration apply softmax again to probability outputs? Is this consistent with the KL objective in Eq. 3?
  2. Why does each sample use the entire batch target vector instead of its own target? Does this affect attribution consistency across different batches?
  3. Could you separately evaluate the contributions of class-guided exploration and displacement-weighted accumulation?
Reviewer_LaiUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new method called Label-Agnostic Attribution for Interpretability (LAAI), that helps us understand which parts of an image influences an AI model's predictions, without requiring to choose a specific class beforehand. Instead of explaining the probability of one particular class, LAAI examines the model's entire probability distribution. It separates class-guided exploration from gradient accumulation of a class-symmetric reverse-KL objective and analyses the objective’s geometry and finite-accumulation properties. The researchers compared LAAI with 11 existing interpretability methods, including Integrated Gradients, SmoothGrad, DeepLIFT, and AttEXplore. Their main experiment used 1000 ImageNet images, three image classification models (Inception-v3, ResNet-50 and VGG16). They mainly evaluated explanations using two tests- Insertion and Deletion. They found that LAAI performed best on insertion, LAAI did not perform best on deletion and LAAI is computationally expensive. The paper also acknowledges that LAAI cannot explain which specific class an image feature supports and also cannot identify every possible change in the model’s predictions.

Critical strengths
  • The class-symmetric objective separates the explanation target from class-guided exploration, providing a clear approach to attribution without requiring a target explanation class.
  • Strong empirical performance across multiple models. LAAI achieves the highest insertion scores among 12 evaluated methods across all six CNN model confidence settings. The results suggest that it effectively identifies image regions that help recover the model's original prediction. The experiments provide evidence that the proposed method is useful and competitive with established techniques, although it performs less well on deletion.
Critical weaknesses
  • There is limited evidence that LAAI genuinely improves interpretability, although LAAI achieves strong insertion scores but it performs worse than several existing methods on deletion as noted in the paper. More importantly, the paper does not establish whether the highlighted image regions are genuinely relevant to the model's predictions. As the central goal of the paper being to improve interpretability, the strong insertion scores alone does not demonstrate that its explanations are more accurate, faithful, or useful.
Questions for authors

It would be helpful if authors could clarify how the KL-divergence attribution objective was implemented in the experiments reported in Tables 1–4, particularly whether the values passed to the logarithm were normalized class probabilities, as required by Equation 3. Additionally, some clarity would be required on which model and data subset were used for the experiments reported in Tables 5 and 10.

3 comments
AuthorsOpenReview ↗

Rebuttal for Author

Thank you for your careful review and valuable feedback. W1: Evidence for feature ranking and prediction relevance.

LAAI's empirical advantage is that it prioritizes image features that recover the original prediction without requiring a user-selected explanation class. Tables 1–3 report the highest insertion scores across all six CNN–confidence settings (Section 4.3). Insertion directly tests this ranking behavior by progressively restoring image features in attribution order and measuring the original prediction score.

This insertion advantage persists when the attribution objective is held fixed. Using the same 1,000 ImageNet images for each of three models and the same reverse-KL objective for all methods, LAAI achieves higher insertion scores in all 12 comparisons against direct gradients, gradient-times-input, and IG50/IG800, with multiplicity-adjusted paired intervals excluding zero. At 800 gradient evaluations per attribution run, insertion improves over IG800 by 0.1515, 0.0986, and 0.1046 on ResNet-50, Inception-v3, and VGG16, respectively; IG800 achieves better deletion on all three models. These comparisons show that exploration and accumulation improve insertion rankings beyond the choice of the KL objective alone.

Q1: Probabilities inside the KL objective in Tables 1–4.

In Tables 1–4, the logarithm in the KL objective takes normalized class probabilities as its input. Given model logits ff, we first compute P=softmax(f)P=\mathrm{softmax}(f) and then evaluate Equation 3: J=KL(Q∥P)=−log⁡C−1C∑jlog⁡Pj.J=\mathrm{KL}(Q\|P)=-\log C-\frac{1}{C}\sum_j\log P_j. The additional softmax in the exploration branch is separate from the KL objective used for accumulation; Appendix C describes this behavior.

Q2: Model and data subset for Tables 5 and 10.

Both Tables 5 and 10 use Inception-v3 on a subset of 1,000 images from the ISA data. They present the same 16-setting M,TM,T grid: Table 5 reports standard insertion/deletion, while Table 10 additionally reports learned-baseline and distributional metrics.

  • Inception-v3 on 1,000 images from “ISA data”. The paper never mentions an ISA dataset. The main experiments use ImageNet. Can you explain the “ISA” here ?

  • At the same default setting of (M = T = 20), Table 5 reports insertion of 0.5948, while Table 11 reports 0.4276 for Inception-v3. If the model and number of images are the same, this gap is not explained?

Thank you for pointing out this ambiguity. We clarify that ISA refers to a subset of 1,000 images sampled from the mini-ImageNet test set, covering 20 classes. This subset was used for the hyperparameter ablation study in Table 5, whereas Table 11 reports the main method comparison on a separate subset of 1,000 ImageNet images.

Although both experiments use Inception-v3 with the same default configuration ((M=T=20)), they are evaluated on different image sets, resulting in different insertion scores.

We will revise the manuscript to explicitly define ISA and clarify the dataset configurations associated with Tables 5 and 11.

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

The paper proposes Label-Agnostic Attribution for Interpretability (LAAI), a method for generating attribution maps without selecting a specific target class. The authors introduce a symmetric distribution-level objective based on reverse KL divergence between the model's predicted probabilities and a uniform distribution. The method uses sampled classes to guide adversarial input exploration while accumulating gradients of the shared objective to estimate feature importance. Experiments are conducted primarily on ImageNet using Inception-v3, ResNet-50, and VGG16, with additional evaluations on ViT-B/16 and comparisons against eleven attribution methods.

Critical strengths

S1. The paper addresses an interesting problem of generating attribution maps without requiring a predefined target class, which could be useful when multiple classes are plausible.

S2. The experimental evaluation includes comparisons against eleven existing attribution methods across multiple model architectures, with improvements in insertion and insertion-deletion gap metrics.

S3. The paper is generally well written and easy to follow. The mathematical formulation and discussion of the limitations of distribution-level attribution are useful.

Critical weaknesses

W1. Limited Methodological Novelty and Insufficient Ablations: The proposed method largely combines existing adversarial gradient-based attribution techniques with a reverse-KL objective. The primary novelty appears to be replacing the class-specific attribution objective with a symmetric distribution-level objective, without introducing a fundamentally new attribution mechanism. The paper does not sufficiently justify why reverse KL is preferable to alternatives such as entropy, forward KL, or logit variance. Controlled comparisons using these objectives within the same exploration framework are necessary to establish the contribution.

W2. Evaluation Does Not Convincingly Establish Attribution Faithfulness: The experimental improvements are primarily concentrated in insertion scores, while deletion performance is consistently weaker than several baselines. For example, on high-confidence Inception-v3 predictions, LAAI obtains a deletion score of 0.1056 compared to 0.0251 for SmoothGrad. More importantly, insertion/deletion metrics alone do not establish whether the highlighted features are semantically relevant or faithful to the model's decision process. The paper lacks ground-truth feature attribution evaluations, controlled synthetic experiments, and statistical significance analysis to substantiate its interpretability claims.

W3. Reproducibility Issues and Computational Inefficiency: The paper acknowledges several inconsistencies between the mathematical formulation and the released implementation, including double-softmax behavior, batch-dependent target selection, uncontrolled randomness, and missing evaluation dependencies.

Questions for authors

See Weakness.

2 comments
AuthorsOpenReview ↗

Rebuttal for Author

Thank you for your careful review and valuable feedback. W1: Contribution and objective comparison.

LAAI separates class-guided exploration from a class-symmetric attribution objective. Sampled classes determine exploration directions, while a shared distribution-level objective determines attribution (Sections 3.2–3.3). Section 3.4 analyzes the objective's geometry and the accumulation using displacements from the original image.

The logit gradient of reverse KL is P−QP-Q, where QQ is uniform. Forward KL differs from negative entropy only by a constant; near uniform predictions, the leading local term of reverse KL is one half of the centered logit variance. To compare these objectives, we used the same 1,000 ImageNet images for each of three models, fixing exploration trajectories, three exploration seeds, M=T=20M=T=20, and ϵ=0.05\epsilon=0.05, and changing only the attribution objective.

The table reports reverse KL minus the alternative; positive ΔINS and negative ΔDEL favor reverse KL. Metrics were averaged over the three seeds within each image before 10,000 paired bootstrap resamples over images. Intervals use Bonferroni correction across 12 comparisons with nominal 95% familywise coverage.

ModelAlternativeΔINS ↑ΔDEL ↓
ResNet-50Forward KL-0.0219 [-0.0269, -0.0170]-0.0122 [-0.0157, -0.0088]
ResNet-50Logit variance+0.1218 [0.1106, 0.1335]-0.0029 [-0.0066, 0.0006]
Inception-v3Forward KL-0.0128 [-0.0154, -0.0103]-0.0048 [-0.0072, -0.0027]
Inception-v3Logit variance+0.1468 [0.1339, 0.1604]-0.0048 [-0.0106, 0.0008]
VGG16Forward KL-0.0099 [-0.0148, -0.0047]-0.0093 [-0.0122, -0.0067]
VGG16Logit variance+0.0783 [0.0697, 0.0872]+0.0008 [-0.0014, 0.0029]

W2: Faithfulness and statistical validation.

Under a fixed KL objective, LAAI improves insertion on all three models in the controlled comparison.

On the same 1,000 ImageNet images for each of three models, we fixed the reverse-KL objective and compared LAAI with direct gradients, gradient-times-input, and IG. The table reports differences against IG800, matched for the gradient budget per attribution run, with multiplicity-adjusted paired confidence intervals.

ModelΔINS ↑ΔDEL ↓
ResNet-50+0.1515 [0.1335, 0.1695]+0.0253 [0.0195, 0.0311]
Inception-v3+0.0986 [0.0815, 0.1158]+0.0467 [0.0381, 0.0557]
VGG16+0.1046 [0.0887, 0.1208]+0.0196 [0.0155, 0.0239]

Δ = LAAI − IG800; higher INS and lower DEL are better. LAAI metrics average three exploration seeds within each image. Brackets show paired bootstrap confidence intervals adjusted for multiple comparisons.

We also completed additive, interaction, and saturating-redundancy tests with known feature dependencies. Ranking performance varies across methods and conditions; complete means and paired comparisons have been added to Appendix G (Tables 30–31) of the revised manuscript, with the additions marked in blue.

W3: Implementation and computational cost.

Section 3.3 and Appendix C now clarify the per-example implementation used in the supplementary experiments. The KL objective follows Equation 3, exploration uses class probabilities directly, and each image has its own sampled target with a fixed random sampling sequence. Appendix C retains the description of the submitted implementation and adds “Implementation for supplementary experiments,” specifying probability computation, target selection, and evaluation settings to connect the mathematical definitions to the supplementary computations.

We have provided the complete evaluation entry point and required dependencies; see Appendix C.2.

The equal-gradient-budget comparison evaluates LAAI's exploration and accumulation. Each exploration step computes two gradients; with M=T=20M=T=20, each attribution run requires 2MT=8002MT=800 gradient evaluations. Against IG800 using the same KL objective and 800 gradient evaluations, LAAI achieves higher insertion scores on all three models; W2 reports the complete comparison for both metrics. Section 4.4 and Table 4 report runtime cost.

We sincerely thank the reviewer for their valuable feedback. We hope that our responses have addressed your concerns. As the rebuttal period is coming to an end, we would be happy to further discuss any remaining questions or concerns. If our responses have resolved your concerns, we would greatly appreciate it if you could reconsider your assessment of our paper in light of our clarifications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.