Research map / Label-efficient and robust learning / Recognition and long-tail
Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes L2P-AHIL, a learning-from-label-proportions method that weights self-training pseudo-labels with a dual entropy-based weight (DEW) at bag and instance levels, improving accuracy especially at large bag sizes.
Problem
Learning from label proportions (LLP) trains instance classifiers from bag-level class proportions; auxiliary pseudo-label instance losses become inaccurate and over-smoothed as bag size grows, degrading performance.
Approach
Adaptively weight each instance's hardened pseudo-label by a product of a bag-level and an instance-level entropy-difference weight (DEW), so only high-confidence pseudo-labels dominate the self-training instance loss jointly optimized with the DLLP bag-level loss.
Key results
- On CIFAR-10 at bag size 256, L2P-AHIL reaches 92.88±0.53% versus 63.10% (ROT) and 48.20% (SoftMatch); its accuracy drops only 2.08 points from bag size 16 to 256.
- On CIFAR-100 at bag size 16 and 32, L2P-AHIL exceeds FLMm by 12.49% and 11.71% respectively.
- Ablation on CIFAR-10 at bag size 256 shows removing both weights drops accuracy to 51.90%, versus 92.88% with both bag-level and instance-level weights.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the instance-level adaptive weight (ablation) | L2P-AHIL · CIFAR-10, bag size 256 | accuracy: 78.03±0.25 -> 92.88±0.53 | Helped |
| Add the bag-level adaptive weight (ablation) | L2P-AHIL · CIFAR-10, bag size 256 | accuracy: 65.54±0.24 -> 92.88±0.53 | Helped |
| Add the instance-level adaptive weight (ablation) | L2P-AHIL · CIFAR-100, bag size 256 | accuracy: 21.59±0.09 -> 23.56±2.13 | Helped |
| Add the bag-level adaptive weight (ablation) | L2P-AHIL · CIFAR-100, bag size 256 | accuracy: 2.89±1.09 -> 23.56±2.13 | Helped |
| Set βb=2 and βi=2 instead of βb=5 and βi=5 (ablation) | L2P-AHIL · CIFAR-100, bag size 128 | accuracy: 72.21 -> 64.81 | Hurt |
| Set βb=40 and βi=40 instead of βb=5 and βi=5 (ablation) | L2P-AHIL · CIFAR-100, bag size 128 | accuracy: 72.21 -> 2.21 | Hurt |
Surprising findings
- Pseudo-label self-training baselines such as SoftMatch collapse at large bag sizes (e.g., 2.12% on CIFAR-100 at bag size 128), despite being strong at small bag sizes.
Limitations
- Hyperparameters of the adaptive weight (beta_b, beta_i) are sensitive and must be tuned for datasets with different class numbers.
- Evaluated only on small, classical benchmark datasets, not larger complex datasets.
- Baseline FLMm results are reproduced from the original publication rather than rerun.
Experimental setup
Datasets: Fashion-MNIST SVHN CIFAR-10 CIFAR-100 MiniImageNet
Benchmarks: CIFAR-10 CIFAR-100 SVHN Fashion-MNIST MiniImageNet
Metrics: classification accuracy
Baselines: LLPFC DLLP LLP-VAT ROT SoftMatch FLMm
Builds on: DLLP weak and strong augmentation SoftMatch ROT
Closest papers by meaning
- Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation (ICLR 2024 · similarity 0.91)
- Rethinking Learning from Label Proportions via Moment Matching (NeurIPS 2026 · similarity 0.90)
- Nearly Optimal Sample Complexity for Learning with Label Proportions (ICML 2025 · similarity 0.88)
- Semi-supervised Multi-label Learning with Balanced Binary Angular Margin Loss (NeurIPS 2024 · similarity 0.87)
- Estimating Noisy Class Posterior with Part-level Labels for Noisy Label Learning (CVPR 2024 · similarity 0.87)
- Multi-Label Learning with Stronger Consistency Guarantees (NeurIPS 2024 · similarity 0.87)
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.