Cross-Domain Few-Shot
Cross-domain few-shot learning addresses the challenging scenario where few-shot tasks at test time come from a different visual or data domain than the tasks seen during meta-training. It tests whether few-shot learning methods truly learn generalizable learning strategies or merely memorize domain-specific features.
The Domain Gap Problem
- Within-Domain: Meta-train on ImageNet classes, meta-test on different ImageNet classes. Feature distributions are similar — the model just needs to handle new categories.
- Cross-Domain: Meta-train on ImageNet, meta-test on medical images, satellite imagery, or industrial inspection data. Feature distributions are fundamentally different — textures, colors, shapes, and visual patterns change entirely.
- Performance Drop: Most meta-learning methods see 15–30% accuracy drops when moving from within-domain to cross-domain evaluation.
BSCD-FSL Benchmark
| Target Domain | Dataset | Description | Visual Gap from ImageNet |
|---|---|---|---|
| Agriculture | CropDisease | Plant disease images | Moderate |
| Satellite | EuroSAT | Satellite land use images | Large |
| Medical | ISIC | Skin lesion dermoscopy | Very large |
| Medical | ChestX | Chest X-ray pathology | Very large |
- Performance degrades as the visual gap from the training domain increases.
- ChestX (most different from ImageNet) shows the worst cross-domain performance.
Why Standard Methods Fail
- Domain-Specific Features: Networks meta-trained on natural images learn features (edges, textures, colors) optimized for that domain. Medical images have entirely different discriminative features.
- Distribution Shift: Pixel distributions, spatial frequencies, and channel statistics differ dramatically across domains.
- Task Structure Mismatch: The "tasks" in different domains have fundamentally different structures — distinguishing dog breeds vs. distinguishing tissue pathologies.
Approaches to Cross-Domain Generalization
- Large Pre-Trained Backbones: Models like CLIP, DINOv2, DeiT trained on massive diverse datasets learn more universal features that transfer better across domains.
- Feature-Wise Transformation Layers (FiLM): Add learnable scaling and shifting parameters that adapt features to new domains without changing the base network.
- Domain-Agnostic Representations: Use adversarial training to learn features that are domain-invariant — a domain discriminator cannot tell which domain the features came from.
- Multi-Source Meta-Training: Train on episodes from multiple diverse source domains simultaneously — increases the diversity of visual experiences.
- Test-Time Adaptation: Fine-tune the feature extractor using the support set from the target domain at test time — adapts representations to the new domain on the fly.
- Self-Supervised Pre-Training: Methods like contrastive learning capture universal visual structure without domain-specific labels.
Current Best Practices
- Start with a large, diverse pre-trained model (CLIP, DINOv2).
- Apply test-time adaptation using the support set.
- Use data augmentation to simulate domain shifts during training.
- Combine metric learning with support set fine-tuning for each new task.
Cross-domain few-shot learning is the true test of meta-learning generalization — methods that only work within a single visual domain are solving a much easier problem than real-world few-shot learning requires.