Research map / Architectures and efficiency / RNNs and normalization
Visualizing Deep Networks by Optimizing with Integrated Gradients
Zhongang Qi, Saeed Khorram, Li Fuxin
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 I-GOS instead of the strongest listed baseline | ImageNet heatmaps with VGG19 · ImageNet, VGG19, 224×224 | deletion and insertion: 0.0482 -> 0.0336 deletion; 0.5139 -> 0.5246 insertion | Improved |
| Use I-GOS instead of the strongest listed baseline | ImageNet heatmaps with ResNet50 · ImageNet, ResNet50, 224×224 | deletion and insertion: 0.0468 -> 0.0420 deletion; 0.5637 -> 0.5846 insertion | Improved |
| Add TV regularization (ablation) | I-GOS · ImageNet, VGG19, 224×224 | insertion: 0.3712 -> 0.5246 | Improved |
| Add TV regularization (ablation) | I-GOS · ImageNet, VGG19, 28×28 | insertion: 0.5181 -> 0.5701 | Improved |
| Add random noise during integrated-gradient computation (ablation) | I-GOS · ImageNet, VGG19, 224×224 | insertion: 0.4194 -> 0.5246 | Improved |
| Add random noise during integrated-gradient computation (ablation) | I-GOS · ImageNet, VGG19, 28×28 | insertion: 0.5634 -> 0.5701 | No clear effect |
| Use backtracking line search instead of a fixed step size (ablation) | I-GOS · ImageNet, VGG19, 224×224 | deletion and insertion: 0.0393 -> 0.0336 deletion; 0.5024 -> 0.5246 insertion | Improved |
| Use backtracking line search instead of a fixed step size (ablation) | I-GOS · ImageNet, VGG19, 28×28 | insertion: 0.5403 -> 0.5701 | Improved |
| Use backtracking line search instead of a fixed step size (ablation) | I-GOS · ImageNet, VGG19, 28×28 | deletion: 0.0906 -> 0.0899 | Hurt |
| Use I-GOS rather than RISE | VGG16 pointing-game evaluation · MSCOCO | pointing game accuracy: 50.71 -> 49.62 | Hurt |
| Use I-GOS rather than RISE | VGG16 pointing-game evaluation · VOC07 | pointing game accuracy: 87.33 -> 85.81 | Hurt |
Closest papers by meaning
- A Novel Visual Interpretability for Deep Neural Networks by Optimizing Activation Maps with Perturbation (AAAI 2021 · similarity 0.85)
- Guided Integrated Gradients: An Adaptive Path Method for Removing Noise (CVPR 2021 · similarity 0.82)
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.82)
- DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.81)
- Great Minds Think Alike: The Universal Convergence Trend of Input Salience (NeurIPS 2024 · similarity 0.81)
- Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · similarity 0.81)
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.
