Research map / Probabilistic and scientific modeling / Molecules and proteins
Manifold-Aligned Guided Integrated Gradients for Reliable Feature Attribution
Soyeon Kim, Seongwoo Lim, Kyowoon Lee, Jaesik Choi
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 |
|---|---|---|---|
| Construct guided attribution path in VAE latent space (MA-GIG) instead of input space (GIG) | Guided IG · Oxford-IIIT Pet | DiffID: 0.3634 -> 0.4637 | Improved |
| Latent-space guided path instead of GIG | Guided IG · ImageNet2012 (500 images) | DiffID: 0.2842 -> 0.3262 (VGG16) | Improved |
| MA-GIG vs IG2 on Pet with VGG16 | MA-GIG · Oxford-IIIT Pet (VGG16) | DiffID: 0.6075 (IG2) vs 0.5913 (MA-GIG) | Hurt |
| MA-GIG vs GIG on deletion AUC | MA-GIG · Oxford-IIIT Pet (VGG16) / ImageNet (5000 images) | Deletion AUC: 0.0333 (GIG) vs 0.0817 (MA-GIG) on Pet VGG16 | Hurt |
| Vary selection fraction q among 0.05, 0.1, 0.2 (ablation) | MA-GIG · Pet, Flower, ImageNet | DiffID: 0.4474 / 0.4360 / 0.4345 (Pet ResNet18) | No clear effect |
| Use Spherical Linear Interpolation (Slerp) instead of linear interpolation (ablation) | MA-GIG · Oxford-IIIT Pet (ResNet18) | DiffID: 0.4637 -> 0.4565 | Mixed |
| Use a domain-matched VAE (MAR, ImageNet-trained) instead of larger SD2 VAE (ablation) | MA-GIG · ImageNet (ResNet18) | DiffID: 0.2627 (MAR) vs 0.2638 (SD2) | No clear effect |
| Use a VAE with lower reconstruction MSE (KD) (ablation) | MA-GIG · ImageNet (ResNet18) | DiffID: MSE 0.00247, DiffID 0.2333 | Hurt |
| Latent-space path vs input-space GIG on completeness residual | MA-GIG · ResNet18, K=200 | completeness residual reduction: 9.8% ImageNet; 33.0% Pet; 30.9% Flower | Improved |
| MA-GIG vs GIG on MPRT sanity check | MA-GIG · ImageNet ResNet18 | MPRT: 0.308 -> 0.625 | Improved |
Closest papers by meaning
- Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution (ICML 2024 · similarity 0.91)
- Beyond Single Path Integrated Gradients for Reliable Input Attribution via Randomized Path Sampling (ICCV 2023 · similarity 0.89)
- Distilled Gradient Aggregation: Purify Features for Input Attribution in the Deep Neural Network (NeurIPS 2022 · similarity 0.88)
- Integrated Decision Gradients: Compute Your Attributions Where the Model Makes Its Decision (AAAI 2024 · similarity 0.85)
- LibraGrad: Balancing Gradient Flow for Universally Better Vision Transformer Attributions (CVPR 2025 · similarity 0.84)
- MFABA: A More Faithful and Accelerated Boundary-Based Attribution Method for Deep Neural Networks (AAAI 2024 · similarity 0.84)
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.
