Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Scaling Few-Shot Learning for the Open World
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.
Introduces few-shot learning with many novel classes (FSL-MNC) on a new ImageNet-MNC benchmark and proposes SHA-Pipeline, which uses 5-way episodic meta-training, distributed parallel training, and hierarchy-aware fine-tuning at test time.
Problem
Standard few-shot learning evaluates only on tens to a few hundred novel classes, while open-world settings involve thousands; episodic meta-training cost grows with way count and accuracy degrades as ways increase.
Approach
Meta-train with a fixed 5-way episodes (since performance gains are nearly independent of way number), parallelize episodes across GPUs by splitting queries, and at meta-test time cluster prototypes into a hierarchy (first-neighbor clustering) and fine-tune the backbone with a CPCC regularizer or a hierarchy-adaptive triplet loss.
Key results
- SHA-Pipeline (CPCC) reaches 49.89% average 1-shot accuracy across 1-2560 ways on ImageNet-MNC, versus 48.61% for P>M>F.
- SHA-Pipeline (CPCC) reaches 68.24% average 5-shot accuracy, 1.16 points above P>M>F.
- The parallel framework achieves a speedup ratio of over 150 at N=2560 ways and runs where the traditional parallel framework runs out of GPU memory for N<160.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hierarchy-aware structured representation learning with CPCC regularization | SHA-Pipeline · ImageNet-MNC, 1-shot, average over ways | mean accuracy (%): 1.28% | Helped |
| Add hierarchy-aware structured representation learning with CPCC regularization | SHA-Pipeline · ImageNet-MNC, 5-shot, average over ways | mean accuracy (%): 1.16% | Helped |
| Use hierarchy triplet loss instead of CPCC regularization (ablation) | SHA-Pipeline · ImageNet-MNC, 1-shot, average over ways | mean accuracy (%): 49.89 -> 49.33 | Mixed |
| Use hierarchy triplet loss instead of CPCC regularization (ablation) | SHA-Pipeline · ImageNet-MNC, 5-shot, average over ways | mean accuracy (%): 68.24 -> 68.57 | Helped |
| Fix meta-training episodes to 5-way rather than matching meta-test way count (ablation) | ProtoNet · ImageNet-MNC, 1-shot | mean accuracy | Helped |
| Keep meta-training episode way count fixed at 5 (ablation) | SHA-Pipeline · ImageNet-MNC meta-training | computation time | Helped |
| Distribute the support set and query subsets across GPUs | SHA-Pipeline meta-training · ImageNet-MNC, 2560-way | speedup: over 150 | Helped |
Surprising findings
- Meta-training performance gains over SimpleShot are almost independent of the number of ways, so 5-way episodes suffice for meta-training in the FSL-MNC setting.
Limitations
- Only image data; text not leveraged.
- Meta-test fine-tuning (50 steps per task) is computationally expensive.
- SHA-Pipeline not fully integrated with meta-training.
Experimental setup
Datasets: ImageNet-1k ImageNet-21K
Benchmarks: ImageNet-MNC
Metrics: mean accuracy
Baselines: ProtoNet ProtoNet-Fix SimpleShot few-shot-baseline P>M>F
Builds on: ProtoNet SimpleShot P>M>F few-shot-baseline
Closest papers by meaning
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.88)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.87)
- Many-Shot In-Context Learning (NeurIPS 2024 · similarity 0.87)
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.87)
- Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · similarity 0.87)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 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.