Research map / Architectures and efficiency / Pruning and efficient tuning
Low-Shot Learning With Imprinted Weights
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 |
|---|---|---|---|
| initialize novel-class classifier weights by imprinting example embeddings instead of random weights | CUB-200-2011 classifier · novel classes, 1-shot | top-1 accuracy: 0.17 -> 21.26 | Helped |
| initialize novel-class classifier weights by imprinting example embeddings instead of random weights | CUB-200-2011 classifier · all classes, 1-shot | top-1 accuracy: 37.36 -> 44.75 | Helped |
| add random cropping and horizontal flipping of novel examples before averaging embeddings (ablation) | Imprinting · CUB-200-2011 novel classes | top-1 accuracy: 21.26 -> 21.40; 39.52 -> 39.35; 45.77 -> 46.35; 49.32 -> 49.80 | No clear effect |
| fine-tune the entire network after imprinting weights (ablation) | CUB-200-2011 classifier · novel classes, 1- and 5-shot | top-1 accuracy: 21.26 -> 18.67; 39.52 -> 46.08 | Helped |
| use imprinted rather than random novel weights to initialize fine-tuning | CUB-200-2011 classifier · novel classes, 5-shot | top-1 accuracy: 34.95 -> 46.08 | Helped |
| use imprinted rather than random novel weights to initialize fine-tuning | CUB-200-2011 transfer learning · new classes, 1-5 shots | top-1 accuracy: 15.90 -> 26.59; 46.21 -> 49.39 | Helped |
| use imprinted rather than random novel weights to initialize fine-tuning | CUB-200-2011 transfer learning · new classes, 20-shot | top-1 accuracy: 71.57 -> 70.07 | Hurt |
| fine-tune all classes jointly rather than imprinting and fine-tuning novel classes | CUB-200-2011 classifier · novel classes, 5-shot | top-1 accuracy: 33.00 vs 46.08 | Hurt |
| average multiple normalized example embeddings into one class template rather than store all examples for nearest-neighbor classification | imprinted classifier · CUB-200-2011 novel classes with multiple examples | top-1 accuracy | Helped |
| use a trainable per-class scale instead of a single global scale factor (ablation) | normalized softmax classifier · CUB-200-2011 | classification accuracy | No clear effect |
| increase embedding dimensionality across the tested settings (ablation) | imprinting model and jointly-trained ConvNet · CUB-200-2011 | classification accuracy | No clear effect |
| train embeddings with softmax loss rather than lifted structured loss | imprinted embedding classifier · CUB-200-2011 novel classes | top-1 accuracy | Helped |
Closest papers by meaning
- Dynamic Few-Shot Visual Learning Without Forgetting (CVPR 2018 · similarity 0.87)
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.85)
- L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement (ICLR 2025 · similarity 0.85)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.85)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.85)
- TrAct: Making First-layer Pre-Activations Trainable (NeurIPS 2024 · 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.