- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper devise Reveal-IG to generate attribution map for the target model. Reveal-IG attributes changes in expected model output along a continuous path of input distributions, moving from broad reference probes toward distributions concentrated around the explained input. Image attribution uses Gaussian probes with changing means and variances; tabular attribution progressively concentrates empirical feature marginals. The method satisfies completeness for the expected responses at its endpoints. Experiments include synthetic diagnostics, ImageNet attribution with ResNet-50 and ViT-B/16, and two tabular regression datasets. Results are particularly strong on sign-sensitive image metrics, while tabular performance is competitive but mixed.
Critical strengths
- A coherent framework. The distributional formulation provides a common view of continuous information revelation across images and tabular data, with a clear completeness property for the quantity being explained.
- Encouraging image results. Reveal-IG substantially improves positive-evidence insertion/deletion and Sensitivity-n over the evaluated baselines on both architectures.
- Useful diagnostics and transparent limitations. The paper includes controlled synthetic examples, endpoint and completeness ablations, confidence intervals, and explicit discussion of runtime and finite-variance endpoints.
Critical weaknesses
(1) The distinction from randomized path attribution is not sufficiently established. SPI (Jeon et al., ICCV 2023) already defines attribution by averaging line integrals over randomized input-space paths. Under Reveal-IG's Gaussian reparameterization, fixing the noise defines an input-space path, and exchanging expectation and integration expresses the image attribution as an expectation of those path attributions. This is a mathematical observation about the proposed formulation, not a claim that the methods are identical: their path distributions and endpoints differ, and the empirical tabular construction has additional scope. However, the related-work distinction between pointwise path averaging and integrating an expected response is insufficient on its own. SPI is neither discussed nor directly compared. The paper should clarify what is conceptually and methodologically new beyond the particular probe family and reveal schedule.
(2) Important practical image baselines are missing. The evaluation includes several relevant IG variants, but omits Grad-CAM and FullGrad. Their different attribution representations do not make them unsuitable for practical insertion/deletion and localization comparisons, particularly on ResNet-50. Completeness is valuable, but does not by itself establish superior empirical usefulness. These comparisons should respect differences in resolution and sign support.
(3) The computational overhead is substantial relative to the demonstrated gains. The paper already reports approximately 1.1 seconds per image versus 165 ms for IG, and approximately 1.5 seconds on Bike Sharing versus 1–3 ms for IG, KernelSHAP, and LIME. The latter overhead is especially concerning because Reveal-IG does not consistently outperform perturbation baselines across tabular metrics. Image settings use 50 integration steps with 10 Monte Carlo samples per step, versus 50 steps or samples for IG and Expected Gradients. Comparisons under equal runtime or model-evaluation budgets, with quality–cost curves, are needed to separate the benefit of the proposed path from additional sampling. The completeness-convergence ablation does not establish attribution-quality gains under matched computation.
(4) The method explains a smoothed prediction rather than necessarily the original prediction. With the default image endpoint σ = .25, completeness concerns the expected output around the input, not f(x) itself. The paper acknowledges this and provides endpoint ablations, but does not sufficiently quantify how often the endpoint changes the target score or class.
(5) Distributional probes do not automatically make intermediate inputs semantically meaningful. Independent Gaussian pixels ignore image structure, while recombining empirical tabular marginals can produce unrealistic combinations of correlated features. These are valid interventional choices, but moving to distribution space does not itself solve the off-distribution problem motivating the method.
(6) Additional evaluation protocols would strengthen the empirical conclusions. Insertion/deletion measures the fixed model's response to perturbed inputs, while RMA measures attribution mass inside object regions. ROAR would provide a complementary test by removing highly ranked features from training and evaluation data and measuring performance after retraining, helping distinguish information removal from sensitivity to distribution shift. The pointing game would test whether the highest-attribution location falls within the target object, complementing RMA's mass-based localization measure. These evaluations would help establish whether the reported gains extend beyond the current protocols; neither should be treated as a complete test of signed attribution faithfulness.
Questions for authors
(1) What is the main difference between Reveal-IG and SPI? Can you compare their results directly?
(2) Can you also compare Reveal-IG with Grad-CAM and FullGrad on the image experiments?
(3) Is Reveal-IG still better when each method is given the same computation time? Can you reduce its cost without losing much accuracy?
(4) How much does adding noise change the model's prediction? Could the better attribution results come mainly from averaging over noisy inputs?
(5) Does the method still work well when input features are strongly related to each other, rather than independent?
(6) Can you add ROAR to check whether removing important features reduces performance after retraining, and the pointing game to check whether the most important pixel lies inside the target object?