Research map / Label-efficient and robust learning / Recognition and long-tail
Training Subset Selection for Weak Supervision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Select weakly labeled examples with the cut statistic instead of using all covered data | Weak supervision pipelines across WRENCH label models, datasets, and end models · WRENCH benchmark | test accuracy: 3.65 points; 61/70 cases improved | Helped |
| Select a subset using the cut statistic | BERT + Data Programming · SemEval test | accuracy: 71.97 -> 91.07 | Helped |
| Select a subset using the cut statistic | RoBERTa + Majority Vote · Yelp test | accuracy: 88.51 -> 95.19 | Helped |
| Select a subset using the cut statistic | BERT + Majority Vote · IMDb test | accuracy: 78.32 -> 81.86 | Helped |
| Select a subset using the cut statistic | RoBERTa + Majority Vote · AGNews test | accuracy: 87.46 -> 88.01 | Mixed |
| Use BioBERT rather than generic BERT representations for cut-statistic selection (ablation) | BERT end model with Majority Vote · CDR test | F1: 59.63 (0.84) -> 61.14 (0.64) | Helped |
| Use BioBERT rather than generic BERT representations for cut-statistic selection (ablation) | Weak-label subsets across CDR label models · CDR | training subset accuracy | Mixed |
| Use SST-fine-tuned BERT rather than generic BERT representations for cut-statistic selection (ablation) | Majority Vote + BERT end model · IMDb test | accuracy: 81.86 -> 87.22 (0.57) | Helped |
| Use a 100-example validation set rather than the full validation set to choose β and the checkpoint (ablation) | Majority Vote weak-supervision pipelines · WRENCH textual datasets | test accuracy | Mixed |
| Use β=0.6 when validation data is unavailable, rather than β=1.0 (ablation) | Weak supervision pipelines · WRENCH label model, dataset, and end-model combinations | median accuracy gain: +1.7 accuracy points | Helped |
| Use the cut statistic rather than entropy scoring to rank weakly labeled examples (ablation) | Five label models with BERT representations · Yelp training subsets | pseudolabel accuracy | Helped |
| Relabel disagreeing examples from their neighbors rather than abstaining on them (ablation) | Cut-statistic subset selection · Weak supervision experiments | end-model performance | No clear effect |
Closest papers by meaning
- Losses over Labels: Weakly Supervised Learning via Direct Loss Construction (AAAI 2023 · similarity 0.90)
- Characterizing the Impacts of Semi-supervised Learning for Weak Supervision (NeurIPS 2023 · similarity 0.90)
- Learning from weak labelers as constraints (ICLR 2025 · similarity 0.90)
- A General Framework for Learning from Weak Supervision (ICML 2024 · similarity 0.89)
- Weak Supervision Performance Evaluation via Partial Identification (NeurIPS 2024 · similarity 0.88)
- Towards a statistical theory of data selection under weak supervision (ICLR 2024 · similarity 0.88)
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.