Research map / Label-efficient and robust learning / Recognition and long-tail
Particle Flow for Learning from Label Proportions
Alain Rakotomamonjy, Maxime Vono, Liva Ralaivola
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 |
|---|---|---|---|
| use FlowLLP anchor particle flow instead of other LLP methods on multiclass tabular datasets | FlowLLP · Multiclass tabular datasets (54 experiments) | number of experiments with best performance: best in 44 of 54 experiments; mean 89.1 vs 88.4 (LLPGAN) | Improved |
| use FlowLLP on binary datasets instead of LLPGAN | FlowLLP · Binary tabular datasets (18 experiments) | balanced accuracy: 83.6 -> 83.7 (best in 10 of 18 vs 7 for LLPGAN) | Improved |
| larger gap in label proportions between classes within bags (ablation) | FlowLLP · Mushrooms (binary), bags of size 50 | anchor correct matching rate | Improved |
| increase diversity of label proportions across bags (ablation) | FlowLLP and LLPGAN · Adult, Mushrooms, Telescope (binary) | balanced accuracy | Improved |
| no label proportion diversity (u = 0) with FlowLLP rather than LLPGAN (ablation) | FlowLLP · binary tabular datasets, controlled bags | balanced accuracy | Hurt |
| add anchor-based loss and particle flow on top of pre-trained ResNet18 with bag loss only (ablation) | end-to-end FlowLLP · CIFAR10, bags of size 50, 64x64 and 224x224 | balanced accuracy | Improved |
| vary anchor loss weight lambda (0, 0.1, 1.0) (ablation) | end-to-end FlowLLP · CIFAR10, bags of size 50 | balanced accuracy | Mixed |
| increase bottleneck dimension of anchors from 50 to 512 (no reduction) (ablation) | end-to-end FlowLLP · CIFAR10 64x64, bags of size 50 | balanced accuracy: 79.6% -> 83.4% | Improved |
| increase number of anchor samples (ablation) | FlowLLP · tabular datasets | balanced accuracy | No clear effect |
| add label regularization in the Wasserstein ground cost (ablation) | FlowLLP · tabular datasets | balanced accuracy | No clear effect |
| increase variance of anchor initialization (ablation) | FlowLLP · tabular datasets | balanced accuracy | Improved |
| add noise to label proportions (ablation) | FlowLLP vs LLPGAN · tabular datasets | balanced accuracy | Hurt |
Closest papers by meaning
- Rethinking Learning from Label Proportions via Moment Matching (NeurIPS 2026 · similarity 0.89)
- Learning from Label Proportions: Bootstrapping Supervised Learners via Belief Propagation (ICLR 2024 · similarity 0.89)
- Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions (CVPR 2025 · similarity 0.89)
- Nearly Optimal Sample Complexity for Learning with Label Proportions (ICML 2025 · similarity 0.87)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.85)
- Predicting Label Distribution from Ternary Labels (NeurIPS 2024 · 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.
