sohamghosh121 / natural-parameter-networks

An implementation of natural parameter networks and its extension to GRUs in PyTorch

Natural Parameter Networks

Natural Parameter Networks are a sampling/variational inference free way for Bayesian Deep Learning, based on the paper [here]{https://arxiv.org/pdf/1611.00448.pdf}.

This repository implements the Gaussian variant of NPN for classification and extends this idea to recurrent architectures.

To run MNIST example

python main-mnist.py

To run LM example

python main-lm.py

You also need to download WikiText 2 data and put it in data/ folder. There should be 3 files train.txt, test.txt, valid.txt.

This has been done for course research project for 10-708 Probablistic Graphical Models at Carnegie Mellon University.

TODO

• Gaussian NPN for regression
• Implement other variants of NPN (Gamma NPN, Poisson NPN)