Research map / Architectures and efficiency / RNNs and normalization
Layer-Wise Relevance Propagation with Conservation Property for ResNet
Seitaro Otsuki, Tsumugi Iida, Félix Doublet, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, Komei Sugiura
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 Relevance Splitting at skip connections, Heat Quantization and z+ rule (full method) vs standard LRP | LRP on ResNet50 · ImageNet val | ID score (%): 1.1 -> 54.5 | Improved |
| propagate relevance through identity-mapping skip connections (ablation) | LRP for ResNet (Ratio-Based Splitting) · ImageNet val | ID score: 0.510 -> 0.545 | Improved |
| use Ratio-Based Splitting instead of Symmetric Splitting (ablation) | LRP for ResNet · ImageNet val | ID score: 0.517 -> 0.545 | Improved |
| use Ratio-Based Splitting without including identity skips (ablation) | LRP for ResNet · ImageNet val | ID score: 0.510 -> 0.510 | No clear effect |
| include identity skips under Symmetric Splitting (ablation) | LRP for ResNet · ImageNet val | Deletion: 0.033 -> 0.036 | Hurt |
| add Heat Quantization (ablation) | LRP for ResNet · ImageNet val | ID score: 0.376 -> 0.510 | Improved |
| use mixture of LRP rules in the proposed method instead of z+ only (ablation) | LRP for ResNet · ImageNet val | ID score (%): 54.5 -> 64.4 | Improved |
| use mixture of rules for baseline LRP instead of epsilon rule (ablation) | Baseline LRP · ImageNet val | ID score (%): 1.1 -> 1.1 | No clear effect |
| proposed method vs Score-CAM on CUB | LRP for ResNet · CUB | ID score (%): 45.7 -> 58.2 | Improved |
Closest papers by meaning
- Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability (NeurIPS 2025 · similarity 0.86)
- AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · similarity 0.84)
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.83)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.83)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.83)
- The Unreasonable Ineffectiveness of the Deeper Layers (ICLR 2025 · similarity 0.83)
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.
