Siamese Network

The defining characteristic of a Siamese network is that its twin subnetworks are identical clones—they share the exact same weights, biases, and hyperparameters. During backpropagation, the gradient updates from both branches are summed and applied to the single shared parameter set. This guarantees that both inputs are processed through the same feature transformation, ensuring that if two inputs are semantically identical, they will be mapped to the exact same point in the embedding space. This weight-tying is what makes the network symmetric and invariant to the order of inputs.

A more sophisticated loss function that uses three inputs simultaneously: an anchor sample, a positive sample (same class as anchor), and a negative sample (different class). The objective is to learn an embedding where the distance between the anchor and positive is smaller than the distance between the anchor and negative by at least a margin α:

• ||f(A) - f(P)||² + α < ||f(A) - f(N)||² This formulation is the foundation of modern facial recognition systems like FaceNet and is superior to contrastive loss because it considers relative distances rather than absolute thresholds, leading to more robust embeddings.

Siamese networks excel at one-shot learning—the ability to make accurate predictions about a new class after seeing only a single example. Because the network learns a similarity function rather than memorizing class boundaries, it can compare a new query image against a single stored exemplar and determine if they match. This is fundamentally different from traditional classifiers that require hundreds of examples per class. Practical applications include:

• Signature verification: Authenticating a signature against a single reference sample.
• Face verification at scale: Matching a face against a database of millions using only one reference photo per identity.
• Drug discovery: Comparing a novel molecule's structure to a single known active compound.

The output of each subnetwork is a dense, low-dimensional vector called an embedding or feature descriptor. The entire training objective is to structure this vector space so that semantic similarity equals spatial proximity. Common techniques to visualize and evaluate this space include:

• t-SNE and UMAP: Dimensionality reduction algorithms that project high-dimensional embeddings to 2D for human inspection.
• Silhouette Score: A metric that measures how well-separated the clusters of different classes are.
• Retrieval mAP: Mean Average Precision used to evaluate how well the embeddings perform on a nearest-neighbor retrieval task. A well-trained embedding space will show distinct, tight clusters for each identity or class, with clear margins between them.

The final layer of a Siamese network computes a distance metric between the two embeddings to produce a similarity score. The choice of metric significantly impacts performance:

• L1 (Manhattan) Distance: Sum of absolute differences. Less sensitive to outliers.
• L2 (Euclidean) Distance: Straight-line distance. Standard choice, penalizes large deviations more heavily.
• Cosine Similarity: Measures the angle between vectors, ignoring magnitude. Preferred when the absolute scale of the embedding is not meaningful.
• Learned Metric: A small feed-forward network or a weighted distance layer can be trained on top of the embeddings to learn an optimal, non-linear similarity function for the specific task.

The mathematical backbone of many Siamese architectures. Cosine similarity measures the cosine of the angle between two embedding vectors, producing a score from -1 to 1. In Siamese networks, the final learned embeddings are often compared using this metric to determine if inputs belong to the same class. It is particularly effective because it is magnitude-invariant, focusing purely on the directional alignment of feature vectors rather than their absolute lengths.

The original loss function designed specifically for Siamese networks. Contrastive loss operates on pairs of data points, pulling similar pairs closer together in the embedding space while pushing dissimilar pairs apart beyond a defined margin. The formula penalizes the network when similar items are too far apart or when dissimilar items are closer than the margin. This creates a well-structured latent space where Euclidean distance directly corresponds to semantic similarity.

A direct architectural evolution of the Siamese network that processes three inputs simultaneously: an anchor, a positive example (same class), and a negative example (different class). The network uses triplet loss to ensure the anchor is closer to the positive than the negative by a specific margin. This triplet structure is the standard for face recognition systems like FaceNet, as it learns a more robust ranking of similarity rather than just a binary match/no-match decision.

The primary application domain for Siamese networks. One-shot learning is the ability to recognize or verify objects from only a single training example per class. A Siamese network achieves this by learning a generalizable similarity function rather than memorizing class-specific features. This is critical for facial verification systems where you cannot retrain a classifier every time a new user is enrolled; the network simply compares the new face against a stored reference image.

A critical metric when Siamese networks are used for deep hashing. In these architectures, the final layer produces a compact binary code rather than a continuous vector. Hamming distance counts the number of bit positions where two binary strings differ. This enables extremely fast, memory-efficient similarity searches using XOR operations on modern hardware. Deep hashing Siamese networks are widely used for large-scale image retrieval where storage and speed are paramount.

A related technique that Siamese networks can be trained to replicate or surpass. Perceptual hashing generates a fingerprint of multimedia content based on its perceptual features, ensuring that visually identical images produce the same hash even after compression or resizing. While traditional pHash uses discrete cosine transforms, a Siamese network can learn a neural perceptual hash that is more robust to specific transformations relevant to a given domain, such as detecting copyrighted material.