Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Contextual Squeeze-and-Excitation for Efficient 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 |
|---|---|---|---|
| Replace instance-based SE with task-level contextual pooling (CaSE) (ablation) | EfficientNetB0 few-shot classifier · VTAB (all), Mahalanobis head | accuracy: 43.6 -> 45.3 | Helped |
| Use CaSE instead of no body adaptation | EfficientNetB0 few-shot classifier · MetaDataset (all) | accuracy: 53.4 -> 69.8 | Helped |
| Replace FiLM generators with CaSE | EfficientNetB0 few-shot classifier · VTAB+MD, average accuracy | accuracy: +1.5% | Helped |
| Use CaSE with reduction 64 rather than FiLM | EfficientNetB0 few-shot classifier · VTAB (all) | accuracy: 44.7 -> 46.2 | Helped |
| Use UpperCaSE rather than LITE | EfficientNetB0 few-shot classifier · VTAB (all) | accuracy: 51.4 -> 58.4 | Helped |
| Use UpperCaSE rather than Big Transfer fine-tuning | VTAB structured · VTAB structured | accuracy: 54.5 -> 39.4 | Hurt |
| Use UpperCaSE instead of FineTuner | ORBIT cross-domain evaluation · ORBIT CLE-VE | frame accuracy: 61.0±2.2 -> 63.0±2.2 | Helped |
| Use UpperCaSE instead of FineTuner | ORBIT cross-domain evaluation · ORBIT CLU-VE | frame accuracy: 48.4±1.9 -> 48.1±1.8 | No clear effect |
| Use a linear output activation rather than sigmoid (ablation) | CaSE · VTAB (all) | accuracy: 56.8 -> 58.4 | Helped |
| Use sigmoid-2 rather than sigmoid output (ablation) | CaSE · MetaDataset (all) | accuracy: 74.9 -> 76.2 | Helped |
| Increase the CaSE hidden-layer count from 2 to 3 (ablation) | UpperCaSE · MetaDataset (all) | accuracy: 76.1 -> 75.5 | Hurt |
| Increase CaSE hidden-layer count from 2 to 4 (ablation) | UpperCaSE · VTAB (all) | accuracy: 58.4 -> 58.0 | Hurt |
Closest papers by meaning
- A Unified Few-Shot Classification Benchmark to Compare Transfer and Meta Learning Approaches (NeurIPS 2021 · similarity 0.87)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.86)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.85)
- Self-Prompt Mechanism for Few-Shot Image Recognition (AAAI 2024 · similarity 0.85)
- Context-Aware Multimodal Pretraining (CVPR 2025 · similarity 0.85)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.85)
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.