Research map / Label-efficient and robust learning / Recognition and long-tail
Editing a classifier by rewriting its prediction rules
Shibani Santurkar, Dimitris Tsipras, Mahalaxmi Elango, David Bau, Antonio Torralba, Aleksander Madry
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 |
|---|---|---|---|
| edit with single synthetic snow-to-road exemplar instead of fine-tuning | VGG16 · Vehicles on snowy roads (Flickr) | error rate | Improved |
| snow-to-road edit side effect on other snow images | VGG16 · ImageNet test images containing snow | accuracy: 92.27 -> 91.05 | No clear effect |
| edit iPod text to blank vs no edit | CLIP-ResNet-50 · Typographic attacks (6 objects) | errors corrected | Improved |
| global fine-tuning instead of editing | CLIP-ResNet-50 · Typographic attacks | accuracy on clean iPod images | Hurt |
| editing vs fine-tuning on non-target classes | VGG16 / ResNet · Synthetic concept-transformation benchmark | % errors corrected: 3.12 -> 32.89 | Improved |
| fine-tuning (local or global) on non-target classes | VGG16 · Synthetic concept-transformation benchmark | % errors corrected | Hurt |
| apply editing constraints to the entire image (-mask) (ablation) | Editing · Synthetic benchmark, ImageNet VGG16 | misclassifications corrected | Improved |
| remove rank-one restriction of the update (-proj) (ablation) | Editing · Synthetic benchmark | misclassifications corrected | Hurt |
| edit deeper layers (ablation) | Editing · Synthetic benchmark, VGG16 | misclassifications corrected | Improved |
| fine-tune deeper layers (local) (ablation) | Local fine-tuning · Synthetic benchmark, VGG16 | non-target class errors corrected | Hurt |
| increase number of exemplars (3 to 10) (ablation) | Editing · Synthetic benchmark, VGG16 | misclassifications corrected | Improved |
| increase number of exemplars (ablation) | Fine-tuning · Synthetic benchmark, VGG16 | misclassifications corrected | Mixed |
Closest papers by meaning
- Editing Boolean Classifiers: A Belief Change Perspective (AAAI 2023 · similarity 0.82)
- Exploring and Leveraging Class Vectors for Classifier Editing (NeurIPS 2025 · similarity 0.80)
- Batchwise Patching of Classifiers (AAAI 2018 · similarity 0.80)
- Exploiting Class Learnability in Noisy Data (AAAI 2019 · similarity 0.79)
- Explainable and Local Correction of Classification Models Using Decision Trees (AAAI 2022 · similarity 0.78)
- Improving Generalization via Meta-Learning on Hard Samples (CVPR 2024 · similarity 0.76)
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.
