Siamese Networks

Siamese networks learn similarity by comparing pairs of examples, useful for verification and few-shot learning. Architecture: Twin networks with shared weights process two inputs, produce embeddings, compare embeddings with distance/similarity metric. Training: Contrastive loss - same-class pairs should be close, different-class pairs far. Triplet loss - anchor closer to positive than negative by margin. Inference: Compare query to each support example, classify by most similar or aggregate similarities. Applications: Face verification (same person?), signature verification, one-shot learning, duplicate detection, image similarity search. Advantages: No retraining for new classes, naturally handles open-set scenarios, learn meaningful similarity metric. Architecture choices: Shared weights (Siamese), different weights for different inputs, various embedding networks (CNNs, transformers). Loss functions: Contrastive loss (pairs), triplet loss (anchor, pos, neg), N-pair loss, InfoNCE. Relationship to metric learning: Siamese instantiates learned distance metric. Modern use: Foundation for contrastive learning, representation learning, still used for verification tasks.