Deep Learning techniques to search for New Physics at LHC

The purpose of this thesis is to apply more recent machine learning algorithms based on neural network architectures in order to discriminate signal from background processes in particle collisions experiments which take place at LHC, in Geneva. First part of this work concerns with neural network architecture and learning algorithm brief description. Then we outline LHC experiments and analysis tools. Finally we introduce our work focusing on the physics of signal process used for our tests and we show our principal results. In particular, we shall confirm that complex neural architectures, namely deep networks, trained on raw kinematics features of particles produced in the process are able to equal or even surpass performances of simpler neural architectures, namely shallow networks, trained on few non linear variables derived from kinematic ones to reduce phase space. This result has a great impact on particle physics research carried on at LHC since it gives a valid alternative analysis tool to classify signals of new physics, especially when non linear features of interest will be yet unknown.