Research map / Label-efficient and robust learning / Recognition and long-tail
Losses over Labels: Weakly Supervised Learning via Direct Loss Construction
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 |
|---|---|---|---|
| Replace pseudolabel-based training with LoL losses and heuristic gradient constraints | weakly supervised text classifiers · WRENCH: AGNews, ChemProt, IMDB, Yelp, YouTube | accuracy | Mixed |
| Add heuristic gradient information to the simple loss (ablation) | LoL-simple · AGNews | accuracy: 82.3 $\pm$ 1.2 -> 83.4 $\pm$ 0.1 | Helped |
| Add heuristic gradient information to the simple loss (ablation) | LoL-simple · ChemProt | accuracy: 53.2 $\pm$ 0.2 -> 52.9 $\pm$ 0.3 | Hurt |
| Add heuristic gradient information to the simple loss (ablation) | LoL-simple · IMDB, Yelp, YouTube | accuracy: IMDB: 80.8 $\pm$ 0.8 -> 81.8 $\pm$ 0.3; Yelp: 74.8 $\pm$ 0.4 -> 75.9 $\pm$ 0.7; YouTube: 93.7 $\pm$ 0.8 -> 94.2 $\pm$ 0. | Helped |
| Use LoL rather than pseudolabel-based baselines with only 100 unlabeled examples | weakly supervised text classifiers · WRENCH: AGNews, ChemProt, IMDB, Yelp, YouTube | accuracy | Helped |
| Use Snorkel-estimated accuracy weights rather than unweighted LoL losses (ablation) | LoL · AGNews, ChemProt, IMDB, Yelp, YouTube | accuracy | Mixed |
| Weight losses by labeler coverage (ablation) | LoL · AGNews and YouTube | accuracy: AGNews: 83.4 $\pm$ 0.1 -> 83.9 $\pm$ 0.2; YouTube: 94.2 $\pm$ 0.7 -> 95.0 $\pm$ 0.6 | Helped |
| Weight losses by labeler coverage (ablation) | LoL · IMDB | accuracy: 81.8 $\pm$ 0.3 -> 75.0 $\pm$ 0.3 | Hurt |
| Replace the squared gradient penalty with an exponential penalty (ablation) | LoL · ChemProt and Yelp | accuracy: ChemProt: 52.9 $\pm$ 0.3 -> 54.2 $\pm$ 0.8; Yelp: 75.9 $\pm$ 0.7 -> 77.1 $\pm$ 2.4 | Helped |
| Replace the squared gradient penalty with a linear penalty (ablation) | LoL · AGNews, IMDB, Yelp, YouTube | accuracy | Hurt |
| Add top-k heuristic gradient constraints to simple losses (ablation) | LoL-simple with image labelers · Animals with Attributes 2 image tasks | accuracy | Helped |
| Use Snorkel-weighted losses (ablation) | LoL on image tasks · raccoon vs. rat | accuracy: 77.9 $\pm$ 4.1 -> 96.4 $\pm$ 0.5 | Helped |
Closest papers by meaning
- Consistent One-vs-All Losses Robust to Misspecification of the Weak Label Transition Model (NeurIPS 2026 · similarity 0.92)
- Learning from weak labelers as constraints (ICLR 2025 · similarity 0.91)
- Loss factorization, weakly supervised learning and label noise robustness (ICML 2016 · similarity 0.91)
- Label Propagation with Weak Supervision (ICLR 2023 · similarity 0.91)
- Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency (ICML 2024 · similarity 0.89)
- Fusing Conditional Submodular GAN and Programmatic Weak Supervision (AAAI 2024 · similarity 0.89)
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.