Research map / Label-efficient and robust learning / Semi-supervised learning
NP-Match: When Neural Processes meet Semi-Supervised Learning
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 KL divergence in NP ELBO with uncertainty-guided skew-geometric JS divergence (ablation) | NP-Match · CIFAR-10, CIFAR-100, STL-10 (9 label settings) | error rate (%): e.g., CIFAR-10 40 labels 5.32 -> 4.91; STL-10 40 labels 14.67 -> 14.20 | Helped |
| use dual form JS_*^{G_alpha_u} instead of primal JS^{G_alpha_u} (ablation) | NP-Match · CIFAR-10, CIFAR-100, STL-10 | error rate (%): comparable; dual slightly better on some settings (e.g., CIFAR-10 250 4.87 vs 4.96) and worse on others (STL-10 1000 5.4 | No clear effect |
| NP-Match (MC-free NP uncertainty) vs MC-dropout UPS | uncertainty estimation · CIFAR-10 and STL-10 test sets | expected UCE (%): CIFAR-10 40: 7.96 -> 7.23; STL-10 40: 17.23 -> 12.45 | Helped |
| NP-Match replacing MC-dropout (UPS) as probabilistic approach | semi-supervised image classification · CIFAR-10 250 labels | error rate (%): 4.96 vs UPS 5.11; CIFAR-10 4000 labels 4.11 vs 4.25; on CIFAR-10 UCE at 4000 labels MC dropout is slightly better (5.82 | Mixed |
| NP-Match vs MC-dropout UPS on ImageNet | ImageNet top-1 error · ImageNet 100 labels per class | top-1 error rate (%): 41.78 vs UPS 42.69 | Helped |
| NP-Match one-pass uncertainty vs MC dropout: inference time grows slowly with number of predictions T | uncertainty estimation time · WRN-28-2 CIFAR-10 and WRN-28-8 CIFAR-100 | time cost (sec): qualitative: UPS time rises quickly with T, NP-Match grows slowly | Helped |
| increase number of sampled latent vectors T below 5 to >=5 (ablation) | NP-Match · hyperparameter exploration CIFAR-10, STL-10 | accuracy: qualitative: performance drops when T<5, little change beyond | Helped |
| increase memory bank length Q (ablation) | NP-Match · hyperparameter exploration CIFAR-10, STL-10 | accuracy: qualitative: performance improves as Q increases | Helped |
| lower uncertainty threshold tau_u (ablation) | NP-Match · hyperparameter exploration CIFAR-10, STL-10 | accuracy: qualitative: lower tau_u leads to worse performance; too large also hurts | Hurt |
Closest papers by meaning
- FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020 · similarity 0.88)
- Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning (AAAI 2024 · similarity 0.87)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.85)
- SemiReward: A General Reward Model for Semi-supervised Learning (ICLR 2024 · similarity 0.85)
- Semi-Supervised Regression with Heteroscedastic Pseudo-Labels (NeurIPS 2025 · similarity 0.85)
- Scaling Up Semi-supervised Learning with Unconstrained Unlabelled Data (AAAI 2024 · similarity 0.85)
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.