This is a brief summary of paper for me to study and organize it, Siamese Neural Networks for One-shot Image Recognition (Koch et al., ICML-WS 2015) I read and studied.

They said whta one-shot learning is:

we may only observe a single example of each possible class before making a prediction about a test instance. This is called one-shot learnig. It is different from zero-shot learning, in which the model cannot look at any example from the target classes.

Their approach to one-shot learning is to use siamese convolutional neural network’s features without any retraing.

They hypothesize that networks which do well at verification should generalize to one-shot classification. i.e. the verfication model learns to identify input pairs according to the probability that they belong to the same class or different class. They used the verification model to evaluate new images, exactly one per novel class, in a pairwise manner against the test images.

ahead of entering siamese convolutional neural network, Let’s see the siamese network:

Koch et al., ICML-WS 2015

The Siamese neural network consists of twin networks which accept distinct inputs but are joined by an energy function(e.g. a loss function) at the top.

The function computes some metric between the highest-level feature representation on each side above image.

Their model’s architecture :

Below show one example of twin networks.

Koch et al., ICML-WS 2015

Reference