Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Channel Importance Matters in 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 |
|---|---|---|---|
| Apply the simple channel-wise feature transformation at test time | Few-shot image classifiers trained on miniImageNet · 19 test-time datasets, 5-way 5-shot | accuracy: 0.5-7.5% | Helped |
| Apply the simple channel-wise feature transformation | ProtoNet features · Plant Disease, binary 5-shot | accuracy: 89.2 -> 93.0 | Helped |
| Replace simple channel transformation with oracle MMC adjustment | ProtoNet + nearest-centroid classifier · Average across 10 datasets, binary 5-shot | accuracy: 83.1 -> 87.7 | Helped |
| Apply simple channel transformation | S2M2 + linear classifier · Average across 10 datasets, binary 5-shot | accuracy: 86.0 -> 87.8 | Helped |
| Replace simple channel transformation with oracle MMC adjustment | S2M2 + linear classifier · Average across 10 datasets, binary 5-shot | accuracy: 87.8 -> 91.7 | Helped |
| Use the transformation on tasks with little or no task-distribution shift | Few-shot classifiers · Same-distribution few-shot evaluation | accuracy | Hurt |
| Use the simple transformation with fine-tuning | Fine-tuning test-time method · miniImageNet, varying shots | accuracy | No clear effect |
| Use the simple transformation with a linear classifier | S2M2 features · miniImageNet, increasing shots | accuracy | Mixed |
| Use Leaky ReLU and extend the transformation to signed features (ablation) | ResNet-12 · 19 test-time datasets, 5-way 5-shot | accuracy | Mixed |
| Use a logarithmic transform without a sufficiently large derivative near zero (ablation) | ProtoNet features · Average across 19 datasets, 5-way 5-shot | accuracy | Hurt |
| Allow a negative derivative in the channel transformation (ablation) | ProtoNet features · Average across 19 datasets, 5-way 5-shot | accuracy | Hurt |
| Use a transformation that loses the negative-second-derivative smoothing property (ablation) | ProtoNet features · Average across 19 datasets, 5-way 5-shot | accuracy | Hurt |
Closest papers by meaning
- Rethinking Generalization in Few-Shot Classification (NeurIPS 2022 · similarity 0.89)
- Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference (CVPR 2022 · similarity 0.89)
- Dense Classification and Implanting for Few-Shot Learning (CVPR 2019 · similarity 0.88)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.87)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.87)
- Enhancing Feature Diversity Boosts Channel-Adaptive Vision Transformers (NeurIPS 2024 · similarity 0.86)
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.