Communication-oriented-Autoencoders---where-Shannon-meets Wiener

The python codes of paper "AUTOENCODERS TRAINED WITH RELEVANT INFORMATION: BLENDING SHANNON AND WIENER’S PERSPECTIVES"

Installation

1. Install an Anaconda environment (recommended).
2. From the environment, update Theano: pip install Theano==0.8.0
3. Install Keras: pip install keras==0.3.3
4. Install Seya: pip install git+https://github.com/EderSantana/Seya.git
5. Install agnez: pip install git+https://github.com/AgnezIO/agnez.git
6. Install seaborn: pip install seaborn
7. Install hdf5: pip install h5py

Demo

1. We upload demos which we train autoencoder architecture with Pri and c-loss funciton.
2. We train four digitals: '0','1','4','5' with four Gaussian prior and visualize the bottleneck layer, which is shown in 'C-Pri-AE-four Gaussian prior demo'. We also add source noise when training, add channel noise when testing and calculate mean square error which is shown in 'C-Pri-AE-ten Gaussian prior demo'. The weights which we have already trained are in 'noise_Gaussian4_pri model' and 'noise_Gaussian10_pri model'

Pri and C-loss

1. The implementation of principle of relevant information(Pri) function shown in the equation(11) is in 'regularization.py'
2. The codes of C-loss objective funciton are in 'objectives.py'