prototypical networks

Prototypical Networks perform few-shot learning by computing class prototypes in learned embedding space. **Core idea**: Examples from same class should cluster together. Represent each class by mean embedding of its examples (prototype). Classify by distance to prototypes. **Algorithm**: Encode support examples → compute prototype per class (mean embedding) → encode query → compute distances to all prototypes → softmax over negative distances for classification. **Distance function**: Typically Euclidean or cosine distance. Euclidean has theoretical justification (Bregman divergences). **Training**: Episodic training matching test-time setup. Sample N-way K-shot tasks from training classes. **Simplicity advantage**: No learned comparison function (unlike Matching Networks), just mean and distance. Fewer parameters, less overfitting. **Extensions**: Task-conditioned prototypes, transductive inference, hierarchical prototypes. **Zero-shot variant**: Use class name embeddings as prototypes. **Performance**: Competitive with more complex meta-learning methods, especially on standard benchmarks. Simple, elegant, widely adopted baseline for few-shot classification.