CCA, DCCA, DCCAE, ConvCCA

This repo contains a pytorch implementation of DCCA, DCCAE, and a CNN-based DCCA.

Please see /data/readme.md for data preparation of Noisy MNIST.

Performance

Reproduce the paper:

Results of ConvCCA:

Differences from the paper:

• Use ReLU instead of Sigmoid/Non-saturating Sigmoid

• Use ADAM optimizer instead of RMSProp/SGD/…

• Add BatchNorm after each layer.

• DCCAE's AE parts are pretrained with ADAM.

References:

• DeepCCA
• On Deep Multi-View Representation Learning