Research map / Learning theory and algorithms / Prediction and decision losses
On the Complexity-Faithfulness Trade-off of Gradient-Based Explanations
Amir Mehrpanah, Matteo Gamba, Kevin Smith, Hossein Azizpour
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 |
|---|---|---|---|
| replace ReLU with smooth parameterization SP(beta=.9) in VanillaGrad (ablation) | CNN · Imagenette (CNN) | EF: .390 -> .202 | Improved |
| apply SmoothGrad post-hoc smoothing (ablation) | ReLU CNN · Imagenette (CNN) | EF: .390 -> .286 | Improved |
| apply SmoothGrad post-hoc smoothing (cost in explanation gap) | ReLU CNN · Imagenette (CNN) | dEF: 0 -> .104 | Hurt |
| apply GradCAM instead of VanillaGrad | ReLU CNN · Imagenette (CNN) | EF: .390 -> .293 | Improved |
| apply SmoothGrad on SP(beta=.9) network instead of ReLU (ablation) | CNN · Imagenette (CNN) | dEF: .104 -> .005 | Improved |
| apply GradCAM instead of VanillaGrad | ResNet50 · ImageNet (pretrained) | dEF: 0 -> .130 | Hurt |
| use ViT-B16 instead of ResNet50 for GradCAM | GradCAM · ImageNet (pretrained) | dEF: .130 -> .041 | Improved |
| use GELU instead of ReLU in ViT-B16 (ablation) | VanillaGrad · Imagenette (ViT-B16 from scratch) | EF: .239 -> .248 | Hurt |
| use GELU instead of ReLU in ViT-B16 (ablation) | GradCAM · Imagenette (ViT-B16 from scratch) | dEF: .051 -> .033 | Improved |
| increase smoothness-parameter beta toward ReLU (ablation) | CNN · Fashion-MNIST, CIFAR10, Imagenette | EF | Hurt |
| add batch normalization (ablation) | CNN · Imagenette 224x224 | spatial power spectrum tail | Hurt |
| add skip connections (ablation) | CNN · Imagenette 224x224 | spatial power spectrum tail | No clear effect |
Closest papers by meaning
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.87)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.87)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.87)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs (ICLR 2025 · similarity 0.86)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.86)
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.
