This project started out as an attempt the replicate the experiment of Lagazio and Marwala with additional economic data about countries.  A significant amount of time was put into finding the data Lagazio and Marwala used as well as finding economic data.  While I did eventually manage to find a data set similar to the one used by Lagazio and Marwala, it used a numerical identification system for countries that made it impossible to relate the data to any other dataset.

In the process of failing to replicate the Lagazio and Marwala experiment I did come across the fantastically useful Matlab library NETLAB.  This library had Bayesian neural networks and ARD already implemented.  After giving up on replicating the Lagazion and Marwala experiment I spent some time playing around with this and trying to characterize of ARD.  These results were incredibly boring: as far as I could tell ARD works exactly as described.  I regret that I did not meticulously document these results, as I would have liked to include them in this report, but I seem to have lost them in Matlab.

Ultimately I became interested in the theory behind ARD.  I choose to write a paper on ARD to further my understanding of the subject.  The resulting paper is, in my opinion, not as good as it should be.  I knowing left key parts of the ARD algorithm (specifically the evidence function, a better explanation of how priors work, and results) and the prose leaves something to be desired.  These faults are mostly due to a lack of time, a result of poor estimation and planning on my part,  Despite these failures I do think that I learned quite a bit from this project- a rough introduction to Bayesian learning, a different way of thinking about neural nets, the problems with arbitrary data, why its important to stick to your topic, etc...

I hope this explanation helps you better understand my project.  Thank you very much.