Research map / Graph learning / GNN architectures
Mutual CRF-GNN for Few-Shot Learning
Check 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 |
|---|---|---|---|
| infer GNN affinity from CRF label-space marginals and mutually update features and affinity | MCGN · miniImageNet and tieredImageNet, transductive 5-way 1-shot and 5-way 5-shot | accuracy: 59.63 ± 0.52 -> 67.32 ± 0.43; 76.34 ± 0.48 -> 83.03 ± 0.54; 63.52 ± 0.53 -> 71.21 ± 0.85; 80.24 ± 0.87 -> 85.98 ± 0.98 | Helped |
| add CRF-inferred affinity and mutual CRF-GNN updating (ablation) | MCGN · CIFAR-FS 5-way 1-shot and 5-way 5-shot, transductive | accuracy: 71.98 ± 0.99 -> 76.45 ± 0.99; 84.22 ± 0.23 -> 88.42 ± 0.23 | Helped |
| add GNN feature aggregation (ablation) | MatchingNet baseline · miniImageNet, tieredImageNet, and CIFAR-FS; 5-way 1-shot and 5-way 5-shot | accuracy: 49.42 ± 0.98 -> 58.93 ± 0.76; 68.20 ± 0.85 -> 76.12 ± 0.94; 53.34 ± 0.78 -> 62.62 ± 0.98; 72.69 ± 0.78 -> 79.64 ± 0.87; | Helped |
| add a single CRF after the backbone (ablation) | MatchingNet baseline · miniImageNet, tieredImageNet, and CIFAR-FS; 5-way 1-shot and 5-way 5-shot | accuracy: 49.42 ± 0.98 -> 53.21 ± 0.76; 68.20 ± 0.85 -> 71.34 ± 0.79; 53.34 ± 0.78 -> 57.43 ± 0.72; 72.69 ± 0.78 -> 76.04 ± 0.73; | Helped |
| make the CRF and GNN branches mutually contribute instead of using separate branches (ablation) | CRF+GNN · miniImageNet, tieredImageNet, and CIFAR-FS; 5-way 1-shot and 5-way 5-shot | accuracy: 60.12 ± 0.57 -> 67.32 ± 0.43; 78.64 ± 0.84 -> 83.03 ± 0.54; 65.43 ± 0.93 -> 71.21 ± 0.85; 82.23 ± 1.02 -> 85.98 ± 0.98; | Helped |
| add CRF marginal-distribution supervision to GNN loss (ablation) | MCGN · miniImageNet 5-way 5-shot | accuracy: about 1% | Helped |
| increase the number of MCGN layers from 0 to 1 (ablation) | MCGN · miniImageNet 5-way 5-shot | accuracy: 72.43% -> 76.34% | Helped |
| increase MCGN depth beyond one layer (ablation) | MCGN · miniImageNet 5-way 5-shot | accuracy | No clear effect |
| use a CRF-inferred class-level affinity rather than EGNN's affinity | MCGN · miniImageNet and tieredImageNet 5-way 1-shot and 5-way 5-shot, transductive | accuracy: 7.69%; 7.23%; 6.69%; 5.74% | Helped |
| use MCGN rather than DPGN | few-shot classification · miniImageNet and tieredImageNet 5-way 1-shot and 5-way 5-shot, transductive | accuracy: 1.2%; 1.8% | Helped |
| use MCGN rather than the Tian et al. distillation baseline | feature-pretrained few-shot learning · miniImageNet and tinyImageNet | accuracy: 4%; 3% | Helped |
| increase the proportion of labeled support samples | MCGN, GNN, and EGNN · miniImageNet semi-supervised 5-way 5-shot | accuracy | Helped |
Closest papers by meaning
- Hybrid Graph Neural Networks for Few-Shot Learning (AAAI 2022 · similarity 0.87)
- Unlocking the Potential of Black-box Pre-trained GNNs for Graph Few-shot Learning (AAAI 2025 · similarity 0.85)
- Preference-driven Knowledge Distillation for Few-shot Node Classification (NeurIPS 2025 · similarity 0.85)
- TINED: GNNs-to-MLPs by Teacher Injection and Dirichlet Energy Distillation (ICML 2025 · similarity 0.84)
- Zero-Shot Generalization of GNNs over Distinct Attribute Domains (ICML 2025 · similarity 0.84)
- Fast Graph Sharpness-Aware Minimization for Enhancing and Accelerating Few-Shot Node Classification (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.