/entropy_dnn

Code for project on relative entropy in deep neural networks

Primary LanguageJupyter NotebookMIT LicenseMIT

entropy_dnn

Code for project on relative entropy in deep neural networks; corresponding paper available at https://arxiv.org/abs/2107.06898. Files include:

  • Gaussian_Feedforward.ipynb -- jupyter notebook that creates and trains feedforward random networks used in the analysis; data written in HDF5 format.

  • Gaussian_Feedforward_Analysis.ipynb -- jupyter notebook that reads HDF5 files created by Gaussian_Feedforward.ipynb and performs analysis (e.g., computes correlation length, generates plots).

  • Relative_Entropy.ipynb -- jupyter notebook that reads HDF5 files created by Gaussian_Feedforward.ipynb and computes the relative entropy or Kullback-Leibler (KL) divergence as a function of depth.

It is recommended to open the .ipynb files with jupyter so that LaTeX expressions in markdown cells are rendered correctly.