To run code, navigate to base directory and run make sgldTest. You can putz around in main.cpp to change the parameters and output file names to generate the charts in the paper.
This was a final project for a PhD-level Monte Carlo Methods course at NYU Courant in spring 2013. The project examines the exactness of the SGLD algorithm against its exact counterpart, Langevin Monte Carlo. I consider a toy probability model similar to the one found in Welling & Whye Teh 2011, the paper that inspired this project. The model is bimodal with a region of low conductance, a shape that is notoriously difficult. The tests I carry out compare the estimates of various probability regions compared to the "true" statistic. Please see the project write up in tex/sgld-lmc-test.pdf for more details.