Wiki · Concept · Last reviewed May 18, 2026

Siamese Networks

Siamese networks are neural architectures that compare two or more inputs by passing them through encoder networks that share weights. Instead of directly classifying a single input, they learn an embedding space where similar inputs are close and dissimilar inputs are far apart.

Definition

A Siamese network uses two identical copies of the same neural network, usually with shared parameters, to encode paired inputs into comparable vectors. The system then measures similarity or distance between those vectors.

The architecture is useful when the task is not "what class is this?" but "are these two things the same, related, or meaningfully similar?"

Mechanism

In a basic setup, input A and input B pass through the same encoder. The encoder produces embeddings. A distance function, similarity score, or learned comparison head then determines whether the pair should be close or far apart.

Training often uses positive pairs and negative pairs. Positive pairs should produce nearby embeddings; negative pairs should produce separated embeddings. This pattern became a foundation for metric learning and later contrastive self-supervised learning.

History

The classic early reference is the 1993 signature-verification work by Bromley, Guyon, LeCun, Sackinger, and Shah. Their system compared signature dynamics by using a Siamese time-delay neural network to decide whether two signatures likely came from the same person.

The deeper idea survived because it is general: learn a representation in which comparison becomes easier.

Uses

Siamese networks appear in face verification, signature verification, image retrieval, duplicate detection, one-shot learning, speaker verification, semantic similarity, and anomaly detection.

They also sit behind later representation-learning families. Contrastive learning, Barlow Twins, VICReg, DINO-style self-distillation, and JEPA all inherit some version of the same question: how can two views, states, or examples be mapped into a useful shared space?

Risk Pattern

Similarity systems can launder bias into distance. If the training pairs reflect social, biometric, or institutional bias, the embedding space can make that bias feel mathematical.

Governance should require clarity about pair construction, negative examples, demographic performance, threshold choice, failure rates, and whether the system is used for high-stakes identification or access control.

Sources

Return to Wiki