Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
A Baseline for Few-Shot Image Classification
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 |
|---|---|---|---|
| use transductive fine-tuning with unlabeled query samples | support-based initialization · Mini-ImageNet 5-way 1-shot (train + val) | accuracy: 58.47 ± 0.66 -> 68.11 ± 0.69 | Helped |
| use transductive fine-tuning with unlabeled query samples | support-based initialization · Tiered-ImageNet 5-way 1-shot (train + val) | accuracy: 67.34 ± 0.69 -> 72.87 ± 0.71 | Helped |
| use transductive fine-tuning with unlabeled query samples | support-based initialization · CIFAR-FS 5-way 1-shot (train + val) | accuracy: 72.14 ± 0.69 -> 78.36 ± 0.70 | Helped |
| use transductive fine-tuning with unlabeled query samples | support-based initialization · FC-100 5-way 1-shot (train + val) | accuracy: 45.08 ± 0.61 -> 50.44 ± 0.68 | Helped |
| use unlabeled-query entropy during fine-tuning (ablation) | fine-tuned classifier and backbone · Mini-ImageNet 5-way 1-shot | accuracy: 59.62 ± 0.66 -> 68.11 ± 0.69 | Helped |
| allow backbone parameters to update during fine-tuning (ablation) | transductive fine-tuning · Mini-ImageNet 5-way 1-shot | accuracy: 58.38 ± 0.66 -> 68.11 ± 0.69 | Helped |
| use logits rather than backbone features as classifier input (ablation) | support-initialized classifier · Mini-ImageNet 5-way 1-shot | accuracy: 64.20 ± 0.65 -> 68.11 ± 0.69 | Helped |
| use logits rather than backbone features as classifier input (ablation) | support-initialized classifier · Mini-ImageNet 5-way 5-shot | accuracy: 81.26 ± 0.45 -> 80.36 ± 0.50 | No clear effect |
| add mixup during pre-training (ablation) | WRN few-shot baseline · Mini-ImageNet 5-way 1-shot | accuracy: 67.06 ± 0.71; about 1% | Helped |
| pre-train on training and validation classes rather than training classes only (ablation) | backbone · Mini-ImageNet 5-way 1-shot, feature-input comparison | accuracy: 62.55 ± 0.65% -> 64.20 ± 0.65% | Helped |
| increase backbone size from WRN-small to WRN (ablation) | transductive fine-tuning · Mini-ImageNet 5-way 1-shot | accuracy: 63.28 ± 0.68% (WRN-small); WRN value reported in Table 1 | Helped |
| use transductive fine-tuning instead of support-based initialization (ablation) | ImageNet-21k few-shot classifier · ImageNet-21k 160-way 1-shot | accuracy: 40.57 ± 0.42 -> 40.46 ± 0.44 | Hurt |
Closest papers by meaning
- Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference (CVPR 2022 · similarity 0.92)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.92)
- Dense Classification and Implanting for Few-Shot Learning (CVPR 2019 · similarity 0.92)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.91)
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.90)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.90)
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.