Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Few-Shot Learning via Embedding Adaptation With Set-to-Set Functions
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 |
|---|---|---|---|
| Adapt support embeddings with a Transformer set-to-set function | ProtoNet embedding model · MiniImageNet 1-shot 5-way, ConvNet | accuracy: 52.61{ $\pm$ 0.20} -> 55.15{ $\pm$ 0.20} | Helped |
| Adapt support embeddings with a Transformer set-to-set function | ProtoNet embedding model · MiniImageNet 1-shot 5-way, ResNet | accuracy: 62.39{ $\pm$ 0.21} -> 66.78{ $\pm$ 0.20} | Helped |
| Use a permutation-sensitive BiLSTM for set adaptation instead of the Transformer (ablation) | FEAT · MiniImageNet 1-shot 5-way, ConvNet | accuracy: 55.15{ $\pm$ 0.20} -> 52.13{ $\pm$ 0.20} | Hurt |
| Replace Transformer set adaptation with DeepSets (ablation) | FEAT · MiniImageNet 1-shot 5-way, ConvNet | accuracy: 55.15{ $\pm$ 0.20} -> 54.41{ $\pm$ 0.20} | Hurt |
| Replace Transformer set adaptation with GCN (ablation) | FEAT · MiniImageNet 1-shot 5-way, ConvNet | accuracy: 55.15{ $\pm$ 0.20} -> 53.25{ $\pm$ 0.20} | Hurt |
| Use DeepSets for interpolation rather than extrapolation across classification ways (ablation) | Set-to-set embedding adaptation · Few-shot classification across 5, 10, 15, and 20 ways | accuracy | Mixed |
| Use GCN for extrapolation rather than interpolation across classification ways (ablation) | Set-to-set embedding adaptation · Few-shot classification across 5, 10, 15, and 20 ways | accuracy | Mixed |
| Replace the Transformer with a deeper or more complex Transformer (ablation) | FEAT · Few-shot classification | accuracy | Hurt |
| Replace max aggregation with sum aggregation in DeepSets (ablation) | DeepSets · Few-shot classification | accuracy | Hurt |
| Adapt support embeddings with Transformer | ProtoNet · TieredImageNet 5-shot 5-way, ResNet | accuracy: 84.03 { $\pm$ 0.16} -> 84.79 { $\pm$ 0.16} | Helped |
| Adapt embeddings for cross-domain few-shot classification | ProtoNet · OfficeHome Clipart to Real World | accuracy: 29.47{ $\pm$0.16} -> 30.89{ $\pm$0.17} | Helped |
| Include unlabeled test instances in the Transformer key and value sets | FEAT · MiniImageNet transductive FSL 1-shot | accuracy: 55.51 -> 57.04 { $\pm$ 0.20} | Helped |
Closest papers by meaning
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.89)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.89)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.89)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.89)
- Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · similarity 0.88)
- FeatWalk: Enhancing Few-Shot Classification through Local View Leveraging (AAAI 2024 · similarity 0.88)
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.