Research map / Architectures and efficiency / RNNs and normalization
Understanding Deep Networks via Extremal Perturbations and Smooth Masks
Ruth Fong, Mandela Patrick, Andrea Vedaldi
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 extremal perturbations instead of RISE | VGG16 pointing game · Pointing game, VOC07 test, all images | pointing game accuracy (%): 86.9 -> 88.0 | Improved |
| use extremal perturbations instead of Grad-CAM | VGG16 pointing game · Pointing game, VOC07 test, all images | pointing game accuracy (%): 86.6 -> 88.0 | Improved |
| use extremal perturbations instead of Grad-CAM | VGG16 pointing game · Pointing game, COCO14 val, all images | pointing game accuracy (%): 54.2 -> 51.5 | Hurt |
| use extremal perturbations instead of Grad-CAM | ResNet50 pointing game · Pointing game, VOC07 test, all images | pointing game accuracy (%): 90.4 -> 88.9 | Hurt |
| use extremal perturbations instead of cMWP | ResNet50 pointing game · Pointing game, COCO14 val, all images | pointing game accuracy (%): 58.5 -> 56.5 | Hurt |
| use extremal perturbations instead of cMWP | VGG16 pointing game · Pointing game, VOC07 test, difficult images | pointing game accuracy (%): 66.5 -> 76.1 | Improved |
| use extremal perturbations instead of center baseline | VGG16 pointing game · Pointing game, COCO14 val, all images | pointing game accuracy (%): 27.8 -> 51.5 | Improved |
| constrain mask area to fixed value rather than penalising it with weights (ablation) | Fong and Vedaldi masks · ImageNet val | monotonic evidence integration (% of images): 98.45 | Improved |
Closest papers by meaning
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.89)
- Perturbation Analysis of Neural Collapse (ICML 2023 · similarity 0.88)
- Deep Perturbation Learning: Enhancing the Network Performance via Image Perturbations (ICML 2023 · similarity 0.88)
- Defense Against Universal Adversarial Perturbations (CVPR 2018 · similarity 0.87)
- Learning perturbation sets for robust machine learning (ICLR 2021 · similarity 0.87)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.86)
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.
