Research map / Label-efficient and robust learning / Recognition and long-tail
Deeply Explain CNN Via Hierarchical Decomposition
Ming‐Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, Philip H. S. Torr
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 average activation instead of average gradient to score channel contribution (ablation) | gAP · ImageNet, T->S5 | PCC: 0.623 -> 0.985 | Improved |
| use average activation instead of maximum activation to score channel contribution (ablation) | gAP · ImageNet, T->S5 | PCC: 0.897 -> 0.985 | Improved |
| use average activation instead of maximum gradient to score channel contribution (ablation) | gAP · PASCAL VOC, S2->S1 | PCC: 0.669 -> 0.893 | Improved |
| use average activation instead of maximum activation (ablation) | gAP · PASCAL VOC, S2->S1 | PCC: 0.897 -> 0.893 | No clear effect |
| replace score-drop ground truth with gAP single backprop for channel contributions | gAP · VGG-16 channel contribution runtime | time per image: 10s -> 50ms, nearly 200x faster | Improved |
| decompose with top-16 features per layer as sparse surrogate model | gAP hierarchical decomposition · ImageNet (VGG-16) | classification accuracy / match rate: top-16 decomposition gives accuracy similar to the original model and consistent predictions on almost all samples | Improved |
| use gAP to propagate importance of selected channels versus individual-based methods when ablating top channels | gAP · ILSVRC val (VGG-16) | accuracy after ablating top channels (lower is better): gAP obtains lower classification accuracy than other individual-based methods | Improved |
Closest papers by meaning
- Interpretable Neural Network Decoupling (ECCV 2020 · similarity 0.86)
- Explaining Convolutional Neural Networks through Attribution-Based Input Sampling and Block-Wise Feature Aggregation (AAAI 2021 · similarity 0.86)
- Towards Better Visualizing the Decision Basis of Networks via Unfold and Conquer Attribution Guidance (AAAI 2023 · similarity 0.86)
- Interpreting Deep Neural Networks with Relative Sectional Propagation by Analyzing Comparative Gradients and Hostile Activations (AAAI 2021 · similarity 0.86)
- Finding Representative Interpretations on Convolutional Neural Networks (ICCV 2021 · similarity 0.86)
- Comparing the Decision-Making Mechanisms by Transformers and CNNs via Explanation Methods (CVPR 2024 · 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.
