Few-Shot Learning Implementation

Build a few-shot learning system for limited data scenarios.

Setup:
- Task: [CLASSIFICATION/DETECTION/NLP]
- N-way K-shot: [N CLASSES, K EXAMPLES]
- Query set size: [QUERIES PER CLASS]
- Meta-training data: [AVAILABLE CLASSES]

Few-shot learning components:
1. Episode construction:
   - Support set sampling
   - Query set sampling
   - Task distribution
2. Metric learning:
   - Prototypical networks
   - Matching networks
   - Relation networks
3. Optimization-based:
   - MAML
   - Reptile
   - Meta-SGD
4. Embedding network:
   - CNN backbone
   - Pretrained features
   - Task-specific adaptation
5. Training:
   - Episodic training
   - Meta-batch construction
   - Validation strategy
6. Evaluation:
   - N-way K-shot accuracy
   - Confidence intervals
   - Cross-domain transfer
7. Practical deployment:
   - Quick adaptation
   - Uncertainty estimation
   - Hybrid approaches

Compare meta-learning vs. transfer learning.