Ezequiel Alvarez (ICAS, Buenos Aires)

Why Bayesian Machine Learning @ HEP?

Bayesian Machine Learning (ML) is a field that has undergone strong theoretical, phenomenological and numerical developments over the last 10-20 years.  Unlike many ML models, instead of using synthetic data to learn, it uses prior information, modeling, and the data itself.  In High Energy Physics, and in particular in those processes where there is still some disagreement between simulations and data, these Bayesian techniques could fill an important gap in New Physics searches and/or in refining our knowledge of the Standard Model.  We will introduce the Bayesian ML framework, review in some detail the above statements, and then show its power in a simplified toy problem inspired by pp -> hh -> bbbb.  We will show how the Bayesian framework provides a unique schema for data-driven methods.  We will discuss the power of statistical tools to leverage simple observations as prior information, such as for instance that a distribution is expected to be continuous, or unimodal. We will show in detail how the method maximizes the exploitation of the multidimensionality of the data and the mutual information contained in it.  We will examine the prospects for applying a Bayesian framework in a real hh -> bbbb analysis.  The general presentation should be useful in understanding how the approach can be applied to a variety of processes, and may sometimes help to find fertile points where searches could be improved by exploiting the full potential of the data.