A standard classifier learns a fixed set of output classes. Add a new face to a face-recognition system, a new product to a catalog, or a new category you've never seen before, and a normal classifier is stuck — it needs retraining with a new output layer.

Siamese Networks sidestep this by never learning to classify at all. Instead, they learn a similarity function: given two inputs, output how similar they are. Once trained, comparing a new, never-before-seen input to a reference is just a forward pass and a distance calculation — no retraining required.

Twin networks, shared weights

A Siamese Network consists of two (or more) identical sub-networks with tied, shared weights. Each copy processes a different input and produces an embedding vector. Because the weights are shared, both inputs are mapped into the same embedding space using the exact same function — this is what makes the resulting distances meaningful and comparable.

The training objective pulls embeddings of similar pairs (two photos of the same person) closer together and pushes embeddings of dissimilar pairs (photos of different people) further apart. The network never explicitly predicts "who is this" — it only ever learns "how far apart should these two embeddings be."

python

Note

A common refinement uses triplets instead of pairs: an anchor, a positive (same class as anchor), and a negative (different class). Triplet loss directly optimizes that the anchor-positive distance be smaller than the anchor-negative distance by at least a margin. This gives the network a relative, rather than absolute, target — often easier to optimize and the basis of FaceNet, one of the most influential face-verification systems.

Face verification: the canonical application

Face verification asks "are these two photos the same person," which is fundamentally a one-shot similarity problem, not a classification problem — you cannot retrain a classifier every time a new employee badge photo is added. A Siamese (or triplet) network trained once on a large face dataset produces an embedding space where same-person photos cluster tightly and different-person photos sit far apart, regardless of whether that specific person's face appeared during training.

Unlocking a phone by face, verifying an ID photo against a selfie, and deduplicating a photo library all use this exact pattern: embed, compare distance, threshold.

Few-shot learning with Siamese networks

The same architecture underlies few-shot classification: given just one or a handful of labeled examples per class, classify a new query by comparing its embedding against the embeddings of the labeled examples and picking the nearest one (nearest-neighbor in embedding space). This is literally how the original Siamese network paper framed one-shot image recognition, and it's conceptually the ancestor of later few-shot approaches like prototypical networks and matching networks.

Siamese networks vs. standard classifiers

Property	Standard classifier	Siamese network
Output	Fixed set of class probabilities	A similarity score / distance between two inputs
New classes	Requires retraining the output layer	Works immediately — no retraining
Training data shape	(input, label)	(input A, input B, same/different) or triplets
Typical use case	Fixed-category classification	Verification, deduplication, few-shot recognition

Note

Siamese and triplet architectures remain the standard for face verification (FaceNet-style systems), signature verification, duplicate detection in search and recommendation, and as a building block inside modern few-shot and retrieval systems. Pure Siamese pair/triplet training has partly been superseded by contrastive pretraining approaches like CLIP and SimCLR for general-purpose embeddings, but the underlying principle — shared weights, embedding distance as the training signal — is identical.

What's next

Siamese networks solve similarity with a fixed pair of examples at inference time. Meta-learning approaches like MAML go further: they learn how to adapt an entire model quickly from just a handful of examples, rather than only comparing embeddings. The next lesson, One-Shot Learning & Meta-Learning, covers that broader family.

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →