This goal of this project is to allow experimentation with different neural network structures and variations on the MCMC sampling procedure.

Inital code based on the MCMC FFN Project by Rohit

The experiments folder contains a bunch of examples of how to train and evaluate models on several datasets. Note that you will need to acquire the raw data and run the processing if you want to change the nature of those datasets.

These scripts will execute the python command line program train_bn_mcmc.py

It expects to be given a training and testing data set, and it expects the data to be a CSV file in which the first number of columns are the numerical input features for the model. And the final number of columns contain the target values.

The value of determines the overall network architecture, and only applies if it is a deep neural network. The value of determines what the activation function will be and you need to choose this depending on the distribution of your target value.

• The above method of describing the neural network structure is cumbersome and inflexible. I plan to make this driven by a single regular expression style syntax that describes the entire architecture.

I am extracting aspects of the MCMC and Metropolis Hastings calculations that are specific to the neural network architecture and embedding them in the specific neural network classes.

This will make the overall MCMC class very abstract/general and I can then easily run multiple architectures side-by-side for comparison.