kamperh / couscous

Siamese neural networks for representation learning using Theano.

Couscous: Siamese Neural Networks for Representation Learning

Overview

This is Theano neural network code, used to train Siamese networks for representation learning. We have used the code here specifically to do representation learning for whole-word speech segments using Siamese convolutional neural networks (CNNs). The code here is the bare neural network code, without feature extraction or evaluation code, or even an example script. We hope to release a separate recipe illustrating the use of Siamese CNNs for representation learning of word segments. For now, at least some of the test functions in couscous/tests/ should be helpful if you want to start to unravel the current code. The name "Couscous" comes from the "cos cos^2" loss used for training Siamese networks.

Dependencies

• Theano and all its dependencies.
• nose

Testing

• Run make test to run unit tests.

Collaborators

• Herman Kamper
• Weiran Wang
• Karen Livescu