Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Dense Classification and Implanting for Few-Shot Learning
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 |
|---|---|---|---|
| Replace global average pooling with dense classification during stage 1 (ablation) | ResNet-12 · miniImageNet 5-way novel-class classification | accuracy: 58.61 ± 0.18 -> 62.53 ± 0.19; 76.40 ± 0.13 -> 78.95 ± 0.13; 80.76 ± 0.11 -> 82.66 ± 0.11 | Helped |
| Add implanting to the dense-classification model (ablation) | ResNet-12 · miniImageNet 5-way 5-shot novel classes | accuracy: 78.95 ± 0.13 -> 79.77 ± 0.19 | Helped |
| Add implanting to the dense-classification model (ablation) | ResNet-12 · FC100 5-way 5-shot novel classes | accuracy: 57.05 ± 0.16 -> 57.63 ± 0.23 | Helped |
| Add implanting to the dense-classification model (ablation) | ResNet-12 · FC100 5-way 10-shot novel classes | accuracy: 61.91 ± 0.16 -> 62.91 ± 0.22 | Helped |
| Widen the last residual block by 16 channels in stage 1 (ablation) | Dense-classification ResNet-12 · miniImageNet 5-way novel classes | accuracy: 62.53 ± 0.19 -> 61.73 ± 0.19; 78.95 ± 0.13 -> 78.25 ± 0.14; 82.66 ± 0.11 -> 82.03 ± 0.12 | Hurt |
| Use dense classification instead of global average pooling with C128F (ablation) | C128F · miniImageNet 5-way novel classes, stage 1 only | accuracy: 54.28 ±0.18 -> 49.84 ±0.18; 71.60 ±0.13 -> 69.64 ±0.15; 76.92 ±0.12 -> 74.61 ±0.13 | Hurt |
| Use dense classification instead of global average pooling with ResNet-12 (ablation) | ResNet-12 · miniImageNet 5-way novel classes, stage 1 only | accuracy: 58.61 ±0.18 -> 61.26 ±0.20; 76.40 ±0.13 -> 79.01 ±0.13; 80.76 ±0.11 -> 83.04 ±0.12 | Helped |
| Use dense classification in stage 1 instead of global average pooling (ablation) | ResNet-12 · miniImageNet 5-way 5-shot, both base and novel classes | accuracy: 56.25 ±0.10 -> 67.53 ±0.10; 54.80 ±0.09 -> 67.78 ±0.10 | Helped |
| Use dense classification at test time for queries instead of global average pooling (ablation) | ResNet-12 trained with dense classification · miniImageNet 5-way 5-shot, both base and novel classes | accuracy: 67.53 ±0.10 -> 67.78 ±0.10; novel classes: 78.55 ±0.13 -> 78.95 ±0.13 | Mixed |
| Use global max-pooling on support and queries instead of average-pooled support and query classification (ablation) | ResNet-12 trained with dense classification · miniImageNet 5-way 5-shot novel classes | accuracy: 79.01 ±0.13 vs 78.95 ±0.13 | No clear effect |
| Use global average pooling rather than global max-pooling for stage 2 support and queries (ablation) | Implanted ResNet-12 · miniImageNet 5-way 5-shot novel classes | accuracy: GMP/GMP 79.03 ± 0.19; GAP/GAP 79.62 ± 0.19 | Mixed |
| Use stage 2 implanting after dense classification (ablation) | ResNet-12 · miniImageNet 5-way 10-shot novel classes | accuracy: 82.66 ± 0.11 -> 83.83 ± 0.16 | Helped |
Closest papers by meaning
- A Baseline for Few-Shot Image Classification (ICLR 2020 · similarity 0.92)
- Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference (CVPR 2022 · similarity 0.91)
- Few-Shot Image Recognition by Predicting Parameters From Activations (CVPR 2018 · similarity 0.91)
- A Closer Look at Few-shot Classification Again (ICML 2023 · similarity 0.91)
- Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · similarity 0.89)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (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.