Research map / Label-efficient and robust learning / Recognition and long-tail
Active Learning Helps Pretrained Models Learn the Intended Task
Alex Tamkin, Dat Pham Nguyen, Salil Deshpande, Jesse Mu, Noah Goodman
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 |
|---|---|---|---|
| uncertainty (least confidence) sampling instead of random acquisition | pretrained BiT-M-R50x1 · Waterbirds | validation accuracy: +11% avg; 5x fewer labels | Improved |
| uncertainty sampling instead of random | RoBERTa-Large · Amazon-WILDS OOD | accuracy: +1% avg; +2.5% worst 10th pct | Improved |
| uncertainty sampling instead of random | BiT-M-R50x1 · iWildCam | accuracy: +9%; 1.8x fewer labels | Improved |
| uncertainty sampling instead of random | BiT-M-R50x1 · Treeperson | accuracy: +2%; 1.6x fewer labels | Improved |
| random initialization instead of pretraining (ablation) | BiT-R50 uncertainty sampling · Waterbirds/iWildCam/Treeperson | accuracy gain from AL | Hurt |
| pretrain on ImageNet-21k instead of ImageNet-1k (ablation) | BiT R50x1 + AL · iWildCam | accuracy vs random | Improved |
| deeper model BiT-M-R101x1 instead of R50x1 (ablation) | BiT-M + AL · Waterbirds/iWildCam/Treeperson | AL gain | No clear effect |
| stop finetuning on training-loss heuristic instead of validation plateau (ablation) | uncertainty sampling BiT · Waterbirds (small experiment) | accuracy: 83.64 -> 85.62 | Improved |
| apply AL on domain far from pretraining (Camelyon17, FMoW) | BiT-M + uncertainty sampling · Camelyon17-WILDS / FMoW-WILDS | accuracy vs random | Hurt |
| more mismatched examples (higher task ambiguity) (ablation) | BiT-M + uncertainty sampling · Waterbirds variants | AL gain over random | Improved |
Closest papers by meaning
- Active Prompt Learning in Vision Language Models (CVPR 2024 · similarity 0.85)
- Contextual Active Model Selection (NeurIPS 2024 · similarity 0.85)
- In-Context Learning Strategies Emerge Rationally (NeurIPS 2025 · similarity 0.84)
- Formal Models of Active Learning from Contrastive Examples (NeurIPS 2025 · similarity 0.84)
- On the Provable Advantage of Unsupervised Pretraining (ICLR 2024 · similarity 0.84)
- On Linear Representations and Pretraining Data Frequency in Language Models (ICLR 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.
