Skip to content

Latest commit

 

History

History
24 lines (13 loc) · 998 Bytes

README.MD

File metadata and controls

24 lines (13 loc) · 998 Bytes

DCCA: Deep Canonical Correlation Analysis

This repository provides a DCCA model, implemented in Keras with tensorflow backend. This repository is developed based on the Keras-theano . For more details, please visit the theano version.

Tips:

After talking with Vahid, we both find that the DCCA loss function based on the automatic gradient is not very stable:

  • Using ReLu instead of sigmoid often cause the gradient exploding.

  • The SGD with momentum works much worse than adam and rmsprop.

  • For each sub-network, the activation function of the last layer should be linear, the second last one should be sigmoid, and all the former ones can be other activation function, e.g., ReLu.

Acknowledgement:

Thanks Vahid for providing the DCCA implementation in Keras-theano and the constructive advises.

Reference:

Galen, A. et. al., Deep Canonical Correlation Analysis, ICML 2013