Research map / Label-efficient and robust learning / Recognition and long-tail
BEE: Metric-Adapted Explanations via Baseline Exploration-Exploitation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BEE, a path-integration explanation method that samples baselines from a learned mixture of baseline distributions, tuned per metric via contextual exploration-exploitation, and selects the best explanation map for the target metric.
Problem
No consensus exists on the best explanation evaluation metric or baseline representation for attribution methods such as Integrated Gradients; different metrics favor different baselines and explanation maps.
Approach
Model the baseline as a random tensor drawn from a learned mixture of baseline types (Constant, Blur, Uniform, Normal, Train Data), with mixture selection learned by a contextual bandit-style exploration-exploitation procedure optimized per metric; integrate over intermediate activations/attention and select the best of T sampled explanation maps.
Key results
- On ResNet101 (IN), fBEE reaches NEG 59.38 and DEL 10.42 vs. best baseline IIA 56.29 and 13.18, respectively.
- On ViT-B, fBEE improves AIC to 66.81 and SIC to 71.45 over the best baseline IIA (62.38 and 68.92).
- Win-rate analysis shows different metrics favor different baselines, e.g., PIC/INS/NEG favor Normal and DEL/POS/SIC favor Uniform.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use fBEE instead of the strongest non-BEE baseline, IIA | ResNet101 explanations · ImageNet validation | NEG: 56.29 -> 59.38 | Helped |
| Use fBEE instead of the strongest non-BEE baseline, IIA | ResNet101 explanations · ImageNet validation | POS: 16.62 -> 13.47 | Helped |
| Use fBEE instead of the strongest non-BEE baseline, IIA | ResNet101 explanations · ImageNet validation | INS: 48.01 -> 51.73 | Helped |
| Use fBEE instead of the strongest non-BEE baseline, IIA | ResNet101 explanations · ImageNet validation | DEL: 13.18 -> 10.42 | Helped |
| Use fBEE instead of the strongest non-BEE baseline, IIA | ViT-Base explanations · ImageNet validation | NEG: 56.01 -> 58.35 | Helped |
| Use fBEE instead of the strongest non-BEE baseline, IIA | ViT-Base explanations · ImageNet validation | POS: 15.19 -> 12.17 | Helped |
| Add inference-time finetuning to pBEE (ablation) | BEE on ResNet101 · ImageNet validation | INS: 51.15 -> 51.73 | Helped |
| Add inference-time finetuning to pBEE (ablation) | BEE on ResNet101 · ImageNet validation | DEL: 11.19 -> 10.42 | Helped |
| Remove contextual input modeling from baseline sampling (ablation) | BEE on ResNet101 · ImageNet validation | NEG: 59.38 -> 58.43 | Hurt |
| Integrate input-image gradients instead of internal representations (ablation) | BEE on ResNet101 · ImageNet validation | NEG: 59.38 -> 51.26 | Hurt |
| Replace BEE with IG using internal representations (ablation) | BEE on ResNet101 · ImageNet validation | NEG: 59.38 -> 55.41 | Hurt |
| Disable inference-time finetuning (ablation) | BEE on ResNet101 · ImageNet validation | POS: 13.47 -> 13.69 | Hurt |
Surprising findings
- Different explanation metrics favor different baseline types, so no single baseline is best across metrics.
Limitations
- Selects the best of a small set of T=8 sampled maps, so the optimum is not guaranteed to be found.
- Finetuning at inference increases runtime due to sequential sampling.
- Evaluated only on image classification with CNN and ViT models on ImageNet.
Experimental setup
Datasets: ImageNet ILSVRC 2012 training subset (5000 examples) ImageNet ILSVRC 2012 validation set
Benchmarks: ImageNet ILSVRC 2012 validation
Metrics: NEG POS INS DEL ADP PIC SIC AIC
Baselines: Grad-CAM Grad-CAM++ Iterated Integrated Attributions FullGrad Ablation-CAM Layer-CAM LIFT-CAM Integrated Gradients Guided IG Blur IG Transformer Attribution Generic Attention Explainability
Builds on: Integrated Gradients Blur IG Expected Gradients Iterated Integrated Attributions Attention Rollout
Closest papers by meaning
- Rethinking Explanation Evaluation Under the Retraining Scheme (AAAI 2026 · similarity 0.86)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (AAAI 2025 · similarity 0.85)
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.85)
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.85)
- Automated Assessment of Fidelity and Interpretability: An Evaluation Framework for Large Language Models’ Explanations (Student Abstract) (AAAI 2024 · similarity 0.85)
- Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 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.