Research map / Probabilistic and scientific modeling / Causal discovery
What is Missing? Explaining Neurons Activated by Absent Concepts
Robin Hesse, Simone Schaub-Meyer, Janina Hesse, Bernt Schiele, Stefan Roth
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 |
|---|---|---|---|
| add non-target (absence) attribution prior to presence debiasing (ablation) | attribution-prior debiasing, X-ResNet-50 · synthetic ISIC, inverse-bias validation | avg accuracy: 0.41 -> 0.59 | Improved |
| add non-target attribution prior (ablation) | attribution-prior debiasing · synthetic ISIC, inverse-bias validation | malignant accuracy: 0.17 -> 0.43 | Improved |
| add non-target attribution prior (ablation) | attribution-prior debiasing · synthetic ISIC, no-bias validation | avg accuracy: 0.77 -> 0.81 | Improved |
| add non-target attribution prior (ablation) | attribution-prior debiasing · synthetic ISIC, training-bias validation | avg accuracy: 0.92 -> 0.89 | Hurt |
| apply any attribution-prior debiasing instead of none (ablation) | X-ResNet-50 · synthetic ISIC, no-bias validation | avg accuracy: 0.51 -> 0.77 | Improved |
| insert least-activating patch instead of random patch | VGG19 · ImageNet-1k, patch 48, 100 images | mean activation: 2.68 -> 0.94 | Improved |
| insert least-activating patch instead of random patch | ResNet-50 · ImageNet-1k, patch 48, 100 images | mean activation: 0.16 -> 0.03 | Improved |
| insert most-activating patch instead of random patch | VGG19 · ImageNet-1k, patch 48, 100 images | mean activation: 2.68 -> 3.88 | Hurt |
| insert prior-work logical NOT patch instead of random patch | VGG19 · ImageNet-1k, patch 48, 100 images | mean activation: 2.68 -> 2.51 | No clear effect |
| increase patch size from 32 to 64 for least-activating patch (ablation) | VGG19 · ImageNet-1k, 100 images | mean activation: 2.14 -> -0.38 | Improved |
| measure least-activating patch in earlier layer layer3[2].conv3 vs layer4[2].conv3 (ablation) | ResNet-50 · ImageNet-1k | mean activation drop: 0.26 -> 0.18 | Mixed |
Closest papers by meaning
- Linear Explanations for Individual Neurons (ICML 2024 · similarity 0.85)
- Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition (ICLR 2024 · similarity 0.84)
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.84)
- CoSy: Evaluating Textual Explanations of Neurons (NeurIPS 2024 · similarity 0.84)
- Evaluating Neuron Explanations: A Unified Framework with Sanity Checks (ICML 2025 · similarity 0.84)
- WWW: A Unified Framework for Explaining What Where and Why of Neural Networks by Interpretation of Neuron Concepts (CVPR 2024 · similarity 0.84)
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.
