Research map / Image generation / Autoregressive and latent generation
Path Choice Matters for Clear Attributions in Path Methods
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 the Concentration Principle for choosing among path-method attribution paths and SAMP, a greedy salient manipulation path interpreter with infinitesimal constraint and momentum strategy, improving Deletion/Insertion on MNIST, CIFAR-10 and ImageNet.
Problem
Path methods for attribution satisfy rigorousness axioms, but distinct path choices give different attributions, so the meaning of attributions is ambiguous and the optimal path is unclear.
Approach
Define a Concentration Principle that prefers path functions maximizing the variance of attributions (concentrating attribution on indispensable features), and approximate the optimal path with a greedy Salient Manipulation Path that uses a Brownian-motion assumption to reduce factorial search to linear; add an L1 step-size bound (IC) and momentum on gradients (MS).
Key results
- SAMP achieves Insertion AUC 1.074 on MNIST and SAMP++ 1.116 on ImageNet, with SAMP++ Deletion 0.145 on ImageNet, outperforming the 12 compared methods on the reported metrics (100 images per dataset).
- Ablation on ImageNet: momentum strategy (MS) gives most of the Deletion/Insertion gain, with MS/IC reaching Deletion 0.145 and Insertion 1.116.
- Sensitivity-N check shows correlation between summed attributions and output rises as the infinitesimal bound eta decreases, then plateaus due to numerical error.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAMP rather than the strongest listed baseline | Path attribution methods · MNIST | Deletion AUC: -0.041 ($\pm$0.14) -> -0.093 ($\pm$0.14) | Helped |
| Use SAMP rather than the strongest listed baseline | Path attribution methods · MNIST | Insertion AUC: 0.808 ($\pm$0.10) -> 1.074 ($\pm$0.18) | Helped |
| Use SAMP rather than the strongest listed baseline | Path attribution methods · CIFAR-10 | Deletion AUC: -0.372 ($\pm$0.54) -> -0.733 ($\pm$0.67) | Helped |
| Use SAMP rather than the strongest listed baseline | Path attribution methods · CIFAR-10 | Insertion AUC: 1.452 ($\pm$0.37) -> 1.458 ($\pm$0.40) | No clear effect |
| Use SAMP rather than the strongest listed baseline | Path attribution methods · ImageNet | Deletion AUC: 0.167 ($\pm$0.13) -> 0.154 ($\pm$0.12) | Helped |
| Use SAMP rather than the strongest listed baseline | Path attribution methods · ImageNet | Insertion AUC: 0.898 ($\pm$0.14) -> 0.984 ($\pm$0.20) | Helped |
| Add momentum strategy (MS) (ablation) | SAMP · ImageNet | Insertion AUC: 0.984 ($\pm$0.195) -> 1.088 ($\pm$0.251) | Helped |
| Add momentum strategy (MS) (ablation) | SAMP · ImageNet | Deletion AUC: 0.154 ($\pm$0.118) -> 0.144 ($\pm$0.115) | Helped |
| Add infinitesimal constraint (IC) (ablation) | SAMP · ImageNet | Deletion AUC: 0.154 ($\pm$0.118) -> 0.159 ($\pm$0.121) | No clear effect |
| Add infinitesimal constraint (IC) (ablation) | SAMP · ImageNet | Insertion AUC: 0.984 ($\pm$0.195) -> 1.056 ($\pm$0.185) | Helped |
| Reduce the step-size upper bound to Bound⁻¹ 1/200 (ablation) | SAMP with IC · ImageNet | Deletion AUC: 0.159 ($\pm$0.121) -> 0.279 ($\pm$0.147) | Hurt |
| Reduce the step-size upper bound to Bound⁻¹ 1/200 (ablation) | SAMP with IC · ImageNet | Insertion AUC: 1.056 ($\pm$0.185) -> 0.939 ($\pm$0.159) | Hurt |
Surprising findings
- Using only the x^T to x^0 direction gives the best Deletion while only x^0 to x^T gives the best Insertion, so the authors combine both directions as a trade-off.
- IC has no significant effect on Deletion/Insertion metrics, since it is designed for completeness rather than metric gains.
Limitations
- Optimality of SAMP as an approximate solver to the Concentration Principle is not theoretically analyzed.
- Requires multiple forward and backward passes per sample, slower than CAM and Grad-CAM.
- Deletion/Insertion are self-comparison metrics with potential information leakage; ground-truth comparison is considered controversial.
Experimental setup
Datasets: MNIST CIFAR-10 ImageNet PASCAL VOC2007 CUB-200-2011
Benchmarks: MNIST CIFAR-10 ImageNet PASCAL VOC2007
Metrics: Deletion AUC Insertion AUC Sensitivity-N Pearson correlation muFidelity pointing accuracy
Baselines: LRP CAM LIME Grad-CAM IG SmoothGrad DeepLIFT RISE XRAI Blur IG Guided IG Deep SHAP
Builds on: Integrated Gradients BlurIG Guided IG RISE
Closest papers by meaning
- From Baselines to Transport Geodesics: Axiomatic Attribution via Optimal Generative Flows (NeurIPS 2026 · similarity 0.85)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.84)
- Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers (CVPR 2025 · similarity 0.83)
- Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model (NeurIPS 2024 · similarity 0.83)
- One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.83)
- Learning Dynamic Evidence Routes for Vision Transformer Probing (NeurIPS 2026 · similarity 0.83)
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.