Hands-On Mathematics for Deep Learning

Siamese neural networks

The Siamese neural network is an architecture that is comprised of two identical neural networks with shared weights, and their parameters are trained to determine the similarity between two data samples using a distance metric on the embeddings. This architecture has proven to be effective for one-shot image classification where the network learns to tell whether or not two images belong to the same class.

In the following diagram, we can see that the network takes in two images and each passes through an identical CNN (f_θ) to generate feature vectors (the embeddings):

Once the embeddings are calculated, we can then ...