Research map / Label-efficient and robust learning / Recognition and long-tail
“Why Not Other Classes?”: Towards Class-Contrastive Back-Propagation Explanations
Yipei Wang, Xiaoqian Wang
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 |
|---|---|---|---|
| weighted contrast instead of original logit back-propagation, keep positive features (Gaussian Blur) | GradCAM · CUB-200 t2 | expected relative probability: 0.406 -> 0.776 | Improved |
| weighted contrast instead of original, keep positive features (Gaussian Blur) | Linear Approximation · CUB-200 t2 | expected relative probability: 0.450 -> 0.749 | Improved |
| weighted contrast instead of original, keep positive features (Gaussian Blur) | GradCAM · Food-101 t2 | expected relative probability: 0.669 -> 0.888 | Improved |
| weighted contrast instead of original, keep negative features (Gaussian Blur) | GradCAM · Food-101 t2 | expected relative probability: 0.317 -> 0.085 | Improved |
| weighted contrast instead of original, keep positive features (Gaussian Blur) | GradCAM · Stanford Cars t1 | expected relative probability: 0.810 -> 0.899 | Improved |
| weighted contrast instead of original, keep positive features (Gaussian Blur) | Linear Approximation · FGVC t1 | expected relative probability: 0.757 -> 0.808 | Improved |
| weighted contrast instead of original, keep positive features (Gaussian Blur) | Linear Approximation · Flower-102 t2 | expected relative probability: 0.598 -> 0.746 | Improved |
| weighted contrast instead of gradient of logit, 1/2/10 PGD iterations | input gradient · ILSVRC2012, 9 CNNs | target probability pt and accuracy | Improved |
| weighted contrast instead of mean contrast | GradCAM · ILSVRC2012 / CUB-200 | explanation quality | Improved |
| weighted contrast instead of max contrast | GradCAM · CUB-200 | explanation quality | Improved |
| change blur baseline to zeros or channel-wise mean (ablation) | weighted GradCAM · CUB-200 t2 | expected relative probability: 0.776 -> 0.762 | No clear effect |
| explain probability instead of logit | back-propagation explanations · CUB-200 | localization vs contrast | Mixed |
Closest papers by meaning
- Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives (NeurIPS 2018 · similarity 0.86)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.85)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.85)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.85)
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 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.
