Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Rethinking Few-shot Image Classification: A Good Embedding is All You Need?
Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B. Tenenbaum, Phillip Isola
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 logistic regression instead of nearest neighbour (ablation) | ResNet-12 embedding · miniImageNet 5-way 5-shot | accuracy (%): 69.96 -> 78.31 | Improved |
| use logistic regression instead of nearest neighbour (ablation) | ResNet-12 embedding · miniImageNet 5-way 1-shot | accuracy (%): 56.29 -> 58.74 | Improved |
| add L2 feature normalization (ablation) | ResNet-12 + LR · miniImageNet 5-way 1-shot | accuracy (%): 58.74 -> 61.56 | Improved |
| add 5x support image augmentation (ablation) | ResNet-12 + LR + L2 · miniImageNet 5-way 1-shot | accuracy (%): 61.56 -> 62.02 | No clear effect |
| add sequential self-distillation of the embedding (ablation) | ResNet-12 + LR + L2 + Aug · miniImageNet 5-way 1-shot | accuracy (%): 62.02 -> 64.82 | Improved |
| add self-distillation (ablation) | Ours · FC100 5-way 5-shot | accuracy (%): 59.13 -> 60.91 | Improved |
| train on meta-train + meta-val (trainval) (ablation) | ResNet-12 embedding · FC100 5-way 5-shot | accuracy (%): 59.1 -> 66.4 | Improved |
| use SEResNet-12 instead of 4-layer ConvNet (ablation) | Ours-distill-trainval · miniImageNet 5-way 1-shot | accuracy (%): 56.64 -> 67.73 | Improved |
| train embedding with multi-way classification instead of multi-task heads (ablation) | ResNet-12 · miniImageNet 5-way 1-shot | accuracy (%): 58.53 -> 62.02 | Improved |
| fine-tune on support images at meta-test | WRN-28-10 embedding · miniImageNet 5-way 1-shot | accuracy (%): 61.1 -> 57.73 | Hurt |
| use self-supervised (MoCo) instead of supervised embedding | ResNet50 · miniImageNet 5-way 1-shot | accuracy (%): 57.56 -> 54.19 | Hurt |
| add distillation to LR classifier (ablation) | ResNet-18 · Meta-Dataset mean accuracy | accuracy (%): 68.02 -> 68.28 | Improved |
Closest papers by meaning
- Meta-Baseline: Exploring Simple Meta-Learning for Few-Shot Learning (ICCV 2021 · similarity 0.93)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.92)
- Few-Shot Image Classification: Just Use a Library of Pre-Trained Feature Extractors and a Simple Classifier (ICCV 2021 · similarity 0.91)
- LibFewShot: A Comprehensive Library for Few-Shot Learning (TPAMI 2023 · similarity 0.91)
- Few-Shot Learning With a Strong Teacher (TPAMI 2022 · similarity 0.91)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (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.
