Bayesian Neural Networks for Classification

We study a model for classification based upon Bayesian statistics. The model, called Bayesian Neural Networks (BNN), is based on a function that is a weighted sum of hyperbolic tangent functions. To illustrate this method, we apply it to the task of separating Supersymmetric (SUSY) events from Standard Model proton-proton events at the LHC. Unlike conventional Neural Networks, the BNN model is an average over networks, which is done by integrating over a high dimensional parameter space. Since integrating over the parameter space is analytically impossible, the BNN method uses Hybrid Markov Chain Monte Carlo techniques to sample the desired probability densities while preserving a high acceptance rate. In this thesis we study the correlation properties of a sequence of Neural Networks. The results of this study are of great importance because validate the strategy currently used by the D0 collaboration, in the search for single top quarks at Fermilab.