Research map / Label-efficient and robust learning / Recognition and long-tail
Lower-Bounded Proper Losses for Weakly Supervised Classification
Shuhei M Yoshida, Takashi Takenouchi, Masashi Sugiyama
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 gLS to backward-corrected cross entropy | Linear model · MNIST | test accuracy: 81.52 ± 1.44 % -> 83.77 ± 0.55 % | Improved |
| Add gLS to backward-corrected cross entropy | MLP · MNIST | test accuracy: 83.09 ± 0.67 % -> 88.63 ± 0.38 % | Improved |
| Add gLS to backward-corrected cross entropy | ResNet-20 · CIFAR-10 | test accuracy: 28.86 ± 2.06 % -> 49.71 ± 3.04 % | Improved |
| Add gLS to backward-corrected cross entropy | WRN-28-2 · CIFAR-10 | test accuracy: 29.57 ± 1.58 % -> 49.98 ± 2.59 % | Improved |
| Tune weight decay for BC (ablation) | Linear model · MNIST | test accuracy: 81.52 ± 1.44 % -> 83.56 ± 0.87 % | Improved |
| Tune weight decay for BC (ablation) | MLP · MNIST | test accuracy: 83.09 ± 0.67 % -> 83.30 ± 1.01 % | No clear effect |
| Tune weight decay for BC (ablation) | ResNet-20 · CIFAR-10 | test accuracy: 28.86 ± 2.06 % -> 29.56 ± 1.49 % | No clear effect |
| Tune weight decay for BC + GA (ablation) | MLP · MNIST | test accuracy: 87.88 ± 1.11 % -> 89.15 ± 0.75 % | Improved |
| Tune weight decay for BC + GA (ablation) | ResNet-20 · CIFAR-10 | test accuracy: 34.39 ± 2.96 % -> 35.36 ± 1.80 % | No clear effect |
| Use gLS with α > 1 rather than α < 1 (ablation) | WRN-28-2 · CIFAR-10 | test accuracy | Improved |
| Vary gLS exponent toward α=1 while remaining in the proper, bounded regime (ablation) | WRN-28-2 · CIFAR-10 | test accuracy | Improved |
| Vary the gLS coefficient k across two orders of magnitude (ablation) | WRN-28-2 · CIFAR-10 | test accuracy | Improved |
Closest papers by meaning
- Loss factorization, weakly supervised learning and label noise robustness (ICML 2016 · similarity 0.91)
- Consistent One-vs-All Losses Robust to Misspecification of the Weak Label Transition Model (NeurIPS 2026 · similarity 0.90)
- Losses over Labels: Weakly Supervised Learning via Direct Loss Construction (AAAI 2023 · similarity 0.89)
- Tight and Fast Bounds for Multi-Label Learning (ICML 2025 · similarity 0.86)
- Provable weak-to-strong generalization via benign overfitting (ICLR 2025 · similarity 0.86)
- Multi-Label Learning with Stronger Consistency Guarantees (NeurIPS 2024 · similarity 0.86)
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.
