Research map / Large language models / AI and society
Attribution Analysis-based Concept Alignment: A Human-in-the-loop Data Debugging Framework
Lei Chai, Lu Qi, Hailong Sun, Jing Zhang, Jingxuan Xu
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 |
|---|---|---|---|
| add priority-score selection (entropy and low confound similarity) of synthetic images (ablation) | AACA app. · Places365 subset, 1500 debug images | average accuracy (%): 86.44 -> 89.92 | Improved |
| add priority-score selection of synthetic images (ablation) | AACA app. · Places365 subset, 2000 debug images | average accuracy (%): 85.25 -> 89.04 | Improved |
| add AACA sel. debug images to training set | MAE baseline · Places365 subset | average accuracy (%): 86.75 -> 89.92 | Improved |
| add AACA app. (unselected) debug images to training set | MAE baseline · Places365 subset | average accuracy (%): 86.75 -> 85.25 | Mixed |
| use human-annotated concepts instead of GPT-4 concepts for prompts (ablation) | AACA sel. · Places365 subset, 1500 debug images | average accuracy (%): 87.42 -> 89.92 | Improved |
| use AACA sel. instead of active-learning selection of search-based images | MAE baseline · Places365 subset, 1500 debug images | average accuracy (%): 88.08 -> 89.92 | Improved |
| swap text-to-image model to FLUX.1 (ablation) | AACA sel. · Places365 subset, 1500 debug images | average accuracy (%): 88.52 -> 89.92 | Improved |
| use AACA sel. with different T2I models (ablation) | AACA sel. · Places365 subset | average accuracy (%): 86.75 -> 89.08 | Improved |
| add AACA debug images to training set | Dinit · Places365 subset concept distribution | Jensen-Shannon divergence: 0.152836 -> 0.084219 | Improved |
| add AACA debug images to training set | Dinit · Human evaluation of concept relevance | Likert score: 3.060 -> 3.520 | Improved |
| add randomly selected search images | MAE baseline · Places365 subset | average accuracy (%): 86.75 -> 86.92 | No clear effect |
Closest papers by meaning
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.84)
- PRIME: Prioritizing Interpretability in Failure Mode Extraction (ICLR 2024 · similarity 0.83)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · similarity 0.83)
- $\texttt{dattri}$: A Library for Efficient Data Attribution (NeurIPS 2024 · similarity 0.83)
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.83)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.83)
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.
