Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Mind the Gap Between Prototypes and Images in Cross-domain Finetuning
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 CoPA, which uses separate transformation heads for prototypes and image instances trained with a symmetric contrastive (CLIP-style) loss, to preserve the prototype-image gap in cross-domain few-shot classification on Meta-Dataset.
Problem
Cross-domain few-shot classification methods that apply one shared transformation to prototypes and image embeddings shrink the natural gap between them, hurting representation clustering and generalization.
Approach
Treat class prototypes like CLIP text prompts and adapt two separate transformation heads (one for prototypes, one for images) with a symmetric cross-entropy contrastive loss, so the prototype-image gap is preserved or enlarged.
Key results
- On Meta-Dataset (train on all datasets), CoPA reaches 76.8% average accuracy across all datasets vs. 75.8% for URL, with the best average rank (3.0) among the compared methods.
- Under ImageNet-only training, CoPA improves unseen-domain average accuracy to 68.7% from URL's 66.6%.
- The gap between prototype and image representations is enlarged from about 0.22 (URL shrinks it) to 1.38 with CoPA in the analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on all datasets, Average Seen | accuracy: 0.7% | Helped |
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on all datasets, Average Unseen | accuracy: 1.5% | Helped |
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on all datasets, Average All | accuracy: 1.0% | Helped |
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on ImageNet only, Average Seen | accuracy: 0.4% | Helped |
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on ImageNet only, Average Unseen | accuracy: 2.1% | Helped |
| Use two separate prototype and image transformations with SCE loss instead of URL’s shared transformation | URL / CoPA · Meta-Dataset, train on ImageNet only, Average All | accuracy: 1.8% | Helped |
| Add TSA modules to CoPA | CoPA · Meta-Dataset, train on all datasets | accuracy | Helped |
| Replace URL’s shared transformation with two transformation heads (ablation) | URL · Meta-Dataset test datasets | accuracy | Mixed |
| Add SCE loss (ablation) | URL · Meta-Dataset test datasets | accuracy | Mixed |
| Replace NCC loss with SCE loss (ablation) | CoPA · Fungi | accuracy: +11.6% | Helped |
| Replace NCC loss with SCE loss (ablation) | CoPA · MSCOCO | accuracy: +18.7% | Helped |
| Enlarge the prototype-image representation gap using CoPA | CoPA · Prototype-image representation analysis | gap: 0.22 to 1.38 | Helped |
Surprising findings
- The shared transformation used by URL numerically shrinks the natural prototype-image gap, while CoPA enlarges it, and validation loss is minimized at the enlarged gap.
- Replacing the SCE loss with the NCC loss in CoPA drops performance drastically (e.g., Fungi down 11.6%, MSCOCO down 18.7%).
Limitations
- Symmetric cross-entropy (contrastive) objective depends on batch size; performance degrades with small batches.
- Only evaluated with ResNet-18 backbone in the presented experiments.
Experimental setup
Datasets: Meta-Dataset ImageNet Omniglot Aircraft Birds Textures QuickDraw Fungi VGG Flower
Benchmarks: Meta-Dataset
Metrics: mean accuracy 95% confidence interval average rank validation loss
Baselines: CNAPs Simple-CNAPs SUR URT Tri-M FLUTE URL TSA TA^2-Net FP-MAML AFP-MAML Finetune
Builds on: URL CLIP NCC loss (prototypical networks) TSA
Closest papers by meaning
- Revisiting Continuity of Image Tokens for Cross-domain Few-shot Learning (ICML 2025 · similarity 0.87)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.86)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.86)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.86)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.86)
- A Closer Look at the CLS Token for Cross-Domain Few-Shot Learning (NeurIPS 2024 · similarity 0.86)
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.