Research map / Label-efficient and robust learning / Recognition and long-tail
A Benchmark for Interpretability Methods in Deep Neural Networks
Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, Been Kim
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 base Gradients, Integrated Gradients, or Guided Backprop rankings instead of random rankings | feature importance evaluation with ROAR · ImageNet, Birdsnap, Food 101 | accuracy | Hurt |
| Use Classic SmoothGrad averaging rather than a single estimate | Gradients and Integrated Gradients · ImageNet, Birdsnap, Food 101 | accuracy | Hurt |
| Use Classic SmoothGrad averaging rather than a single estimate | Guided Backprop · ImageNet, Birdsnap, Food 101 | accuracy | No clear effect |
| Add squared noisy-estimate aggregation (SmoothGrad-Squared) | base feature importance estimators · ImageNet, Birdsnap, Food 101 | accuracy | Improved |
| Aggregate noisy estimates by variance (VarGrad) | base feature importance estimators · ImageNet, Birdsnap, Food 101 | accuracy | Improved |
| Square a single estimate before ranking, without ensembling (ablation) | Gradients, Integrated Gradients, and Guided Backprop · ImageNet, Birdsnap, Food 101 | accuracy: 8.43% ± 0.97 | Improved |
| Replace removal with keeping the most important pixels (KAR) (ablation) | feature importance estimators · ImageNet, Birdsnap, Food 101 | accuracy | No clear effect |
| Retrain on the modified training data rather than score modified inputs without retraining (ablation) | ROAR evaluation · ImageNet | accuracy: 0.5% versus 63.53% | Improved |
| Add VarGrad aggregation rather than using a single estimate (ablation) | Guided Backprop · Birdsnap, ROAR 90.0% removed | accuracy: 59.76 -> 5.20 | Improved |
| Add SmoothGrad-Squared aggregation rather than using a single estimate (ablation) | Guided Backprop · Birdsnap, ROAR 90.0% removed | accuracy: 59.76 -> 5.21 | Improved |
| Add SmoothGrad-Squared aggregation rather than using a single estimate (ablation) | Gradients · Food 101, ROAR 90.0% removed | accuracy: 78.36 -> 8.34 | Improved |
| Add VarGrad aggregation rather than using a single estimate (ablation) | Gradients · ImageNet, ROAR 90.0% removed | accuracy: 66.75 -> 10.41 | Improved |
Closest papers by meaning
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.87)
- Testing Semantic Importance via Betting (NeurIPS 2024 · similarity 0.86)
- Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · similarity 0.86)
- Understanding Impacts of High-Order Loss Approximations and Features in Deep Learning Interpretation (ICML 2019 · similarity 0.86)
- Not All Samples Are Created Equal: Deep Learning with Importance Sampling (ICML 2018 · similarity 0.85)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · 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.
