Research map / Label-efficient and robust learning / Recognition and long-tail
Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Penalize importance assigned to spurious image patches | VGG16 on ISIC · ISIC all | F1: 0.56 -> 0.60 | Helped |
| Penalize gradients with RRR | VGG16 on ISIC · ISIC all | F1: 0.56 -> {0.44} | Hurt |
| Exclude patch-containing images from training | VGG16 on ISIC · ISIC no patches | F1: 0.56 -> 0.59 | Helped |
| Penalize single-pixel contributions to encourage shape-based predictions | CNN on ColorMNIST · ColorMNIST | accuracy: 0.2 $\pm$ 0.2 -> 31.0 $\pm$ 2.3 | Helped |
| Penalize squared gradients with RRR | CNN on ColorMNIST · ColorMNIST | accuracy: 0.2 $\pm$ 0.2 -> 0.2 $\pm$ 0.1 | No clear effect |
| Penalize Expected Gradients | CNN on ColorMNIST · ColorMNIST | accuracy: 0.2 $\pm$ 0.2 -> 10.0 $\pm$ 0.1 | No clear effect |
| Penalize importance assigned to the gray decoy patch | CNN on DecoyMNIST · DecoyMNIST | accuracy: 60.1 $\pm$ 5.1 -> 97.2 $\pm$ 0.8 | Helped |
| Hide race from the model | two-layer DNN on COMPAS · COMPAS | WCR (Black): 0.47$\pm$ 0.03 -> 0.44$\pm$0.02 | Helped |
| Encourage explanations to reflect race and its interactions | two-layer DNN on COMPAS · COMPAS | WCR (Black): 0.47$\pm$ 0.03 -> 0.39$\pm$0.04 | Helped |
| Encourage explanations to reflect race and its interactions | two-layer DNN on COMPAS · COMPAS | test accuracy: 67.8$\pm$1.0 -> 68.8$\pm$0.3 | No clear effect |
| Penalize explanations for injected random indicator words | LSTM on SST · SST Random words unbiased test | accuracy: 56.6 $\pm$ 5.8 -> 75.4 $\pm$ 0.9 | Helped |
| Penalize explanations for biased articles | LSTM on SST · SST biased articles unbiased test | accuracy: 57.8 $\pm$ 0.8 -> 68.2 $\pm$ 0.8 | Helped |
Closest papers by meaning
- Learning with Explanation Constraints (NeurIPS 2023 · similarity 0.88)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.86)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.86)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.85)
- Towards Modeling Uncertainties of Self-Explaining Neural Networks via Conformal Prediction (AAAI 2024 · similarity 0.85)
- Explanation Bottleneck Models (AAAI 2025 · similarity 0.85)
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.