Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Few-Shot Classification With Feature Map Reconstruction Networks
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 |
|---|---|---|---|
| Use feature-map reconstruction with closed-form ridge regression for class scoring | Few-shot classification · Fine-grained benchmarks: cropped CUB, raw CUB, Aircraft, meta-iNat, tiered meta-iNat | accuracy: FRN is superior across the board; 2-7 point jump in accuracy | Helped |
| Use FRN with pre-training and episodic fine-tuning (ablation) | FRN · mini-ImageNet | 5-shot accuracy: 78.01 -> 82.83 | Helped |
| Use pre-training without episodic fine-tuning (ablation) | FRN · mini-ImageNet | 5-shot accuracy: 78.01 -> 75.11 | Hurt |
| Use global-average-pooling pre-training rather than FRN pre-training, then fine-tune episodically (ablation) | FRN · mini-ImageNet | 5-shot accuracy: 78.01 -> 76.28 | Hurt |
| Train using 5-shot rather than 1-shot episodes (ablation) | FRN · 5-way 1-shot evaluation on fine-grained benchmarks | accuracy: CUB Conv-4: 68.27±0.23 -> 73.48±0.21; Aircraft Conv-4: 50.25±0.21 -> 53.20±0.21 | Helped |
| Remove the auxiliary loss (ablation) | FRN · Cropped CUB | accuracy: Conv-4: 68.27±0.23 -> 68.38±0.23; 88.43±0.13 -> 88.13±0.13 | No clear effect |
| Learn lambda rather than fixing it (ablation) | FRN · Cropped CUB, Conv-4 | accuracy: 66.44±0.23 -> 68.27±0.23 | Helped |
| Learn rho rather than fixing it (ablation) | FRN · Cropped CUB, Conv-4 | accuracy: 67.34±0.23 -> 68.27±0.23 | Helped |
| Learn lambda and rho rather than fixing both (ablation) | FRN · Cropped CUB | accuracy: Conv-4: 66.19±0.23 -> 68.27±0.23; ResNet-12: 82.80±0.19 -> 82.62±0.19 | Mixed |
| Use FRN pre-training followed by fine-tuning rather than training from scratch (ablation) | FRN · CUB raw images | 5-shot accuracy: 92.92±0.10 -> 93.59±0.09 | Helped |
| Use FRN rather than DeepEMD for inference reconstruction | Few-shot classifier with ResNet-12 · mini-ImageNet, 5-way 5-shot meta-test latency | latency (ms): >800,000 -> 79 | Helped |
| Use FRN rather than CTX | Few-shot classifier · CUB training, 10x10 feature maps, 5-shot | memory (MB): 9693 -> 6727 | Helped |
Closest papers by meaning
- Bi-directional Feature Reconstruction Network for Fine-Grained Few-Shot Image Classification (AAAI 2023 · similarity 0.91)
- A Baseline for Few-Shot Image Classification (ICLR 2020 · similarity 0.90)
- CAD: Co-Adapting Discriminative Features for Improved Few-Shot Classification (CVPR 2022 · similarity 0.90)
- Dense Classification and Implanting for Few-Shot Learning (CVPR 2019 · similarity 0.90)
- A Closer Look at Few-shot Classification (ICLR 2019 · similarity 0.90)
- Simple Semantic-Aided Few-Shot Learning (CVPR 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.