Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Intra-Block Fusion and Cross-Scale Attention modules on top of fine-tuned ResNet50/Swin-S backbones (pre-trained on ImageNet with supervised learning or DINO) to set new few-shot results on in-domain and cross-domain benchmarks.
Problem
Few-shot classification with few labeled target samples, where fine-tuning pre-trained CNN/Transformer backbones suffers from feature drift in intermediate blocks and scale inconsistency between local and global features, especially under cross-domain shift.
Approach
Fuses all intra-block layer features via a 1x1 convolution (IBF) and applies cross-scale attention across the four block-level feature maps (CSA) so each scale's features are enriched by others, then fine-tunes the backbone on episodic tasks with nearest-centroid classification; for cross-domain tasks, adds a meta-testing stage that fine-tunes on support-set augmentations.
Key results
- On FC100 5-way 5-shot, Ours (Swin-S) reaches 86.02% vs. 82.98% for P>M>F (Swin-S), more than three points better.
- On Meta-Dataset, the method achieves 81.61% average accuracy with RN50 vs. 77.61% for P>M>F, a 3.77-point gain in average accuracy.
- Ablation on Swin-S shows IBF+CSA raises miniImageNet 5-way 5-shot from 96.58 to 99.04 (combined), and cross-scale attention beats self-attention by up to 4.32 points (RN50 FC100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the proposed feature-strengthening and feature-weighting method | few-shot classification with Swin-S · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot and 5-way 5-shot | accuracy: 84.41 -> 87.06; 92.16 -> 93.62; 95.26 -> 97.03; 98.20 -> 99.04; 69.92 -> 73.07; 82.98 -> 86.02 | Helped |
| Add the proposed method | RN50 few-shot classifier · Meta-Dataset, 8 in-domain datasets, average | accuracy: 77.61 -> 81.61 | Helped |
| Add the proposed method | RN50 few-shot classifier · Meta-Dataset, INet in-domain, average | accuracy: 73.25 -> 77.02 | Helped |
| Add IBF without CSA (ablation) | RN50 few-shot classifier · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot | accuracy: 77.98; 86.36; 61.01 | Helped |
| Add CSA without IBF (ablation) | RN50 few-shot classifier · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot | accuracy: 78.04; 87.08; 63.24 | Helped |
| Add CSA to IBF (ablation) | RN50 few-shot classifier · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot | accuracy: 77.98 -> 78.62; 86.36 -> 88.78; 61.01 -> 64.03 | Helped |
| Add IBF to CSA (ablation) | RN50 few-shot classifier · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot | accuracy: 78.04 -> 78.62; 87.08 -> 88.78; 63.24 -> 64.03 | Helped |
| Replace local-scale features with cross-scale features (ablation) | RN50 few-shot classifier · CIFAR-FS and miniImageNet, 5-way 1-shot | accuracy: 77.73 -> 78.62; 85.93 -> 88.78 | Helped |
| Replace local-scale features with cross-scale features (ablation) | Swin-S few-shot classifier · CIFAR-FS, miniImageNet, and FC100, 5-way 1-shot and 5-way 5-shot | accuracy: 85.97 -> 87.06; 92.58 -> 93.62; 95.51 -> 97.03; 97.81 -> 99.04; 70.28 -> 73.07; 84.19 -> 86.02 | Helped |
Limitations
- Evaluated only on image classification; the authors note segmentation and crowd counting remain future work.
- Cross-domain fine-tuning depends on carefully selected per-domain learning rates chosen from a small grid using extra episodes.
Experimental setup
Datasets: ImageNet-1k miniImageNet CIFAR-FS FC100 CDFSL Meta-Dataset
Benchmarks: miniImageNet CIFAR-FS FC100 CDFSL Meta-Dataset
Metrics: accuracy
Baselines: P>M>F MetaOpt-SVM EASY HCTransformers ProtoNet Finetune DeepCluster2 SUR URL ITA CTX
Builds on: PMF (P>M>F) CrossViT Crossformer++ DenseNet DINO
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 0.91)
- Meta-Exploiting Frequency Prior for Cross-Domain Few-Shot Learning (NeurIPS 2024 · similarity 0.90)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.90)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.89)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.89)
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.89)
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.