Dichotomize and Generalize: PAC-Bayesian Binary Activated Deep Neural Networks

This repository contains an implementation of PBGNet (PAC-Bayesian Binary Gradient Network) and all related experiments presented in "Dichotomize and Generalize: PAC-Bayesian Binary Activated Deep Neural Networks" by Letarte, Germain, Guedj and Laviolette, accepted at NeurIPS 2019.

Requirements

• Python 3.6
• Numpy 1.14.3
• Pytorch 1.2.0
• Poutyne 1.2
• Scikit-learn 0.20.3
• Pandas 0.23.0
• Click 6.7

Launching

To reproduce the experiment presented in Section 6 of the paper, run:

python launch.py

To launch a single learning experiment with custom options, use experiment.py. Here is an example:

python experiment.py -d mnist17 -n pbgnet --experiment-name my_exp --sample-size 50 --hidden-size 25

For all possible options and their description, see python experiment.py --help.

BiBTeX

@inproceedings{letarte2019dichotomize,
  title={Dichotomize and generalize: Pac-bayesian binary activated deep neural networks},
  author={Letarte, Ga{\"e}l and Germain, Pascal and Guedj, Benjamin and Laviolette, Fran{\c{c}}ois},
  booktitle={Advances in Neural Information Processing Systems},
  pages={6869--6879},
  year={2019}
}