Research map / Probabilistic and scientific modeling / Molecules and proteins
Look at the Variance! Efficient Black-box Explanations with Sobol-based Sensitivity Analysis
Thomas FEL, Remi Cadene, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, Thomas Serre
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use total Sobol indices to include higher-order interactions instead of first-order indices alone (ablation) | Sobol attribution · ImageNet validation | Deletion: 0.170 -> 0.121 | Improved |
| use signed total Sobol indices instead of standard total Sobol indices (ablation) | Sobol attribution · VOC07 Test and COCO14 Val | Pointing Game accuracy: 87.7 -> 89.8; 75.3 -> 82.4; 54.6 -> 55.2; 49.6 -> 49.9; 89.6 -> 86.7; 80.2 -> 73.0; 57.3 -> 57.5; 52.7 -> 53.1 | Mixed |
| use signed total Sobol indices instead of standard total Sobol indices (ablation) | Sobol attribution · ImageNet validation | Deletion: 0.121 -> 0.145; 0.109 -> 0.114; 0.104 -> 0.147; 0.107 -> 0.141 | Hurt |
| reduce the number of forwards from the default 3936 to 492 (ablation) | Sobol attribution on ResNet50V2 · ImageNet validation | Deletion: 0.121 -> 0.151 | Hurt |
| reduce the number of forwards from 3936 to 2460 (ablation) | Sobol attribution on ResNet50V2 · ImageNet validation | Deletion: 0.121 -> 0.121 | No clear effect |
| lower the word-mask threshold to 0.05 so fewer words are removed (ablation) | Sobol signed on BERT · IMDB word deletion | Word deletion: 0.598 -> 0.527 | Improved |
| use Sobol rather than RISE | Black-box attribution · ImageNet validation, EfficientNet convergence | Spearman rank correlation: 0.65 -> 0.8 after 1,000 forwards; 0.73 -> 0.97 with 7,000 forwards | Improved |
| use Sobol rather than RISE | Black-box attribution · Pointing Game benchmarks | Pointing Game accuracy: RISE to Sobol: VOC VGG16 86.9 to 87.7, ResNet50 86.4 to 89.6; COCO VGG16 50.8 to 54.6, ResNet50 54.7 to 57.3 | Improved |
| use Sobol rather than RISE | Black-box attribution · ImageNet validation | Deletion: 0.127 -> 0.121; 0.121 -> 0.109; 0.119 -> 0.104; 0.114 -> 0.107 | Improved |
| use Sobol signed rather than standard Sobol | Black-box attribution · IMDB word deletion | Word deletion: 0.662 -> 0.598; 0.523 -> 0.461 | Improved |
| use Sobol signed rather than standard Sobol | Black-box attribution · ImageNet validation | Insertion: 0.370 -> 0.258; 0.313 -> 0.290; 0.309 -> 0.204; 0.331 -> 0.211 | Hurt |
Closest papers by meaning
- Better Understanding Differences in Attribution Methods via Systematic Evaluations (TPAMI 2024 · similarity 0.85)
- Towards More Robust Interpretation via Local Gradient Alignment (AAAI 2023 · similarity 0.85)
- The Edge of Orthogonality: A Simple View of What Makes BYOL Tick (ICML 2023 · similarity 0.84)
- One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.84)
- On Gradient-like Explanation under a Black-box Setting: When Black-box Explanations Become as Good as White-box (ICML 2024 · similarity 0.83)
- The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric (ICLR 2024 · 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.
