CNN MNIST Siamese network

This example shows a simple example of the DAG interface for a MNIST Siamese network. This network is a simple extension of the original MNIST network with two branches connected to a Contrastive loss [1].

The training is done on the original MNIST data and learns an embedding to a 2D space which is simple to visualize.

Install module

The simplest way how to install the module is to run:

vl_contrib install siamese-mnist
vl_contrib setup siamese-mnist
edit siamese_mnist_example

Example Code

An example tutorial on how to run the training and evaluate a MNIST Siamese network is in the MATLAB Live Script siamese_mnist_example.mlx. It shows how to train and cut the Siamese network to parts in order to obtain the final feature embedding.

The final embedding should look similar to this:

Contents of the package

• cnn_mnist_siames.m Network and training definition script.
• matlab/vl_create_siamese.m Helper function to create a Siamese network.
• matlab/vl_nncontrloss.m Implementation of the Contrastive loss.

[1] Hadsell, Raia, Sumit Chopra, and Yann LeCun. "Dimensionality reduction by learning an invariant mapping" In Proc. of CVPR 2006.