Research map / Learning theory and algorithms / Prediction and decision losses
Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Integrated Decision Gradients (IDG), a path-integral attribution method that weights each gradient by the path derivative of the output logit so that the decision region dominates, plus an adaptive sampling scheme for non-uniform Riemann steps.
Problem
Integrated Gradients suffers from the saturation effect: gradients from path regions where the output logit barely changes dominate and produce noisy, poor attributions for image classifiers.
Approach
Scale each path gradient by the importance factor dF/dalpha (derivative of output with respect to path position), so gradients in the decision region are amplified and saturated ones suppressed; combine with adaptive non-uniform sampling placed where the logit grows.
Key results
- IDG with adaptive sampling reaches RISE deletion 0.108 on ResNet101 vs 0.181 for IG, and achieves the best score across all 12 metric-model combinations.
- On ResNet101 ablation, IDG with uniform sampling (0.546 insertion) beats IG with adaptive sampling (0.532), supporting that the importance factor, not sampling, addresses saturation.
- IDG with 50 steps and adaptive sampling matches IDG with 600 uniform steps in qualitative comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the strongest baseline with IDG | ImageNet attribution methods, three models · ImageNet val, R101/R152/RNXT | AIC: 0.675 -> 0.701; 0.686 -> 0.718; 0.654 -> 0.730 | Helped |
| replace the strongest baseline with IDG | ImageNet attribution methods, three models · ImageNet val, R101/R152/RNXT | SIC: 0.609 -> 0.638; 0.619 -> 0.659; 0.554 -> 0.620 | Helped |
| replace the strongest baseline with IDG | ImageNet attribution methods, three models · ImageNet val, R101/R152/RNXT | Insertion: 0.561 -> 0.592; 0.577 -> 0.615; 0.307 -> 0.324 | Helped |
| replace the strongest baseline with IDG | ImageNet attribution methods, three models · ImageNet val, R101/R152/RNXT | Deletion: 0.148 -> 0.108; 0.148 -> 0.118; 0.078 -> 0.068 | Helped |
| add adaptive sampling to IG (ablation) | IG, ImageNet ResNet101 · ImageNet, ResNet101 | Insertion: 0.513 -> 0.532 | Helped |
| add adaptive sampling to IDG (ablation) | IDG, ImageNet ResNet101 · ImageNet, ResNet101 | Insertion: 0.546 -> 0.586 | Helped |
| add adaptive sampling to IG (ablation) | IG, ImageNet ResNet101 · ImageNet, ResNet101 | Deletion: 0.167 -> 0.146 | Helped |
| add adaptive sampling to IDG (ablation) | IDG, ImageNet ResNet101 · ImageNet, ResNet101 | Deletion: 0.135 -> 0.110 | Helped |
Surprising findings
- Adaptive sampling alone applied to IG yields only minor gains, indicating the importance factor is the main source of improvement.
Limitations
- Axiomatic guarantees hold only under monotonic path and non-decreasing model assumptions.
- Evaluated only on image classifiers (three ImageNet CNNs) with a single black baseline.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet
Metrics: AIC AUC SIC AUC RISE insertion AUC RISE deletion AUC
Baselines: IG Left-IG Guided IG AGI
Builds on: Integrated Gradients Left-IG Guided IG Adversarial Gradient Integration RISE XRAI
Closest papers by meaning
- Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution (ICML 2024 · similarity 0.86)
- Iterative Search Attribution for Deep Neural Networks (ICML 2024 · similarity 0.85)
- Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model (NeurIPS 2024 · similarity 0.84)
- FRInGe: Distribution-Space Integrated Gradients with Fisher–Rao Geometry (NeurIPS 2026 · similarity 0.83)
- MFABA: A More Faithful and Accelerated Boundary-Based Attribution Method for Deep Neural Networks (AAAI 2024 · similarity 0.83)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.