Bayesian Neural Neworks in Data-Intensive High Energy Physics Applications
Abstract
Bayesian Neural Networks (BNNs) are applied to data-intensive High Energy Physics (HEP) applications for classification and regression. Neural networks are non-linear functions that can be used to model (in principle) any mapping of N continuous real variables to M real variables. Where traditional neural networks use optimization techniques to find an optimum set of neural network parameters, ω0, BNNs assign a probability density to every set of network parameters, ωk, in the parameter space. However, BNNs are significantly more computationally intensive to construct than neural networks. The goal of this work is to develop efficient implementations of the training of BNNs on Graphical Processing Units (GPUs).
Our preliminary studies with a GPU indicate that speed improvements of at least 80 are already possible with relatively modest optimization. We are therefore confident that the outcome of the proposed work could be extremely far-reaching once it becomes possible to fit complicated multivariate functions in minutes rather than in hours or days.