Research map / Probabilistic and scientific modeling / Molecules and proteins
Making Sense of Dependence: Efficient Black-box Explanations Using Dependence Measure
Paul Novello, Thomas FEL, David Vigouroux
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 HSIC attribution instead of RISE | black-box attribution · ImageNet val, ResNet50 | Deletion: 0.114 -> 0.106 | Improved |
| use HSIC attribution instead of RISE | black-box attribution · ImageNet val, ResNet50 | Insertion: 0.554 -> 0.470 | Hurt |
| use accurate variant (p=1536) instead of efficient (p=764) | HSIC attribution · ImageNet val, ResNet50 | Insertion: 0.470 -> 0.481 | Improved |
| use HSIC eff. instead of RISE for execution time | black-box attribution · ImageNet, ResNet50 | Exec. time (s): 8.427 -> 0.956 | Improved |
| use HSIC eff. instead of Sobol for execution time | black-box attribution · ImageNet, ResNet50 | Exec. time (s): 5.254 -> 0.956 | Improved |
| use HSIC instead of D-RISE on object detection | YOLOv4 · COCO val | Insertion: 0.634 -> 0.658 | Improved |
| use HSIC instead of D-RISE on object detection | YOLOv4 · COCO val | muFidelity: 0.442 -> 0.568 | Improved |
| use HSIC instead of Kernel Shap on object detection | YOLOv4 · COCO val | Deletion: 0.070 -> 0.088 | Hurt |
| use HSIC instead of D-RISE for runtime | YOLOv4 · COCO val | Exec. time (s): 155 -> 34 | Improved |
| use HSIC instead of Kernel Shap for runtime | YOLOv4 · COCO val | Exec. time (s): 192 -> 34 | Improved |
| use HSIC eff. instead of Integrated Gradients (white-box) for Deletion | attribution · ImageNet val, ResNet50 | Deletion: 0.138 -> 0.106 | Improved |
| use HSIC eff. instead of GradCAM++ (white-box) for Insertion | attribution · ImageNet val, ResNet50 | Insertion: 0.497 -> 0.470 | Hurt |
Closest papers by meaning
- Self-Supervised Learning with Kernel Dependence Maximization (NeurIPS 2021 · similarity 0.87)
- What You See is What You Classify: Black Box Attributions (NeurIPS 2022 · similarity 0.85)
- Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · similarity 0.84)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.83)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.83)
- On Gradient-like Explanation under a Black-box Setting: When Black-box Explanations Become as Good as White-box (ICML 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.
