This is an implementation of Correlational Neural Network (CorrNet) described in the following paper :
Sarath Chandar, Mitesh M Khapra, Hugo Larochelle, Balaraman Ravindran. Correlational Neural Networks. To appear in Neural Computation, 2015.
Please site this paper if you are using this code for any of your publications.
To run the representation learning code, you need Python and Theano.
To run the MNIST example, you need scikit-learn.
Refer section 5 in the paper for details about the two-view setup for MNIST and the transfer learning experiment. First you should download the dataset from this link and extract all the files to some directory say MNIST_DIR. You also need a target directory where the models will be saved, say TGT_DIR.
In terminal, go to mnistExample folder.
To create the dataset for training, run the following command:
$ python create_data.py MNIST_DIR/
Next, to train the CorrNet, run the following command.
$ python train_corrnet.py MNIST_DIR/ TGT_DIR/
To project the data to the learnt space, run the following command.
$ python project_corrnet.py MNIST_DIR/ TGT_DIR/
To evaluate the learnt model for transfer learning task, run the following command.
$ python evaluate.py tl TGT_DIR/
With batch_size=100, training_epochs=50, l_rate=0.01, optimization="rmsprop", tied=True, n_hidden=50, lambda=2, hidden_activation=sigmoid, output_activation=sigmoid, loss_fn = "squarrederror", you should get 77.05% accuracy for view1 to view2 and 78.81% accuracy for view2 to view1.
To compute sum correlation in the projected space, run the following command.
$ python evaluate.py corr TGT_DIR/
With the same configuration as above, you should get 42.57 as test correlation.
Scripts to run the bilingual word representation learning algorithm will be made available soon.