The LHC particle accelerator at CERN is probing the elementary building blocks of matter at energies never seen in laboratory conditions before. In the process of providing new insights in to the Standard Model describing the current understanding of physics governing the behaviour of particles, the accelerator is challenging the algorithms and techniques used in storing the collected data, rebuilding the collected collision events from the detector signal and analysing the data. For this end many state of the art methods are being developed by the scientist working in the LHC experiments in order to gain as much knowledge from the unique data collected from these particle collisions. The decade starting from 2010 can be in many respects considered as the deep learning revolution where a family of machine learning algorithms collectively called deep neural networks had significant breakthroughs driven by advances in hardware used to train these algorithms. During this period many achievements previously only seen in the realm of science fiction became reality as the deep neural networks began driving cars, images and videos could be enhanced with super resolution in real time and improvements in automated translation tools lowered the barriers in communication between people. These results have given the field of deep learning a significant momentum and lead to the methods spreading across academic disciplines as well as different industries. In this thesis the recent advances of deep learning are applied into the realm of particle physics using the data collected by the CMS experiment at the LHC at CERN. First topic presented considers the task of rebuilding the flight paths of charged particles called tracks inside the detector using the measurements made by the Tracker sub-detector in the heart of the CMS. The conditions present inside the detector during particle collisions demand for advanced algorithms able to be both fast and precise. The project in this thesis looks at estimating the quality of the reconstructed tracks and reject tracks that look like they are a result of mistakes made by the reconstruction algorithms, purifying the reconstructed dataset from false signals. Previously the task has been done initially by cut based selections determined by physicists and later by another machine learning algorithm known as the boosted decision tree. Here the first application of deep neural networks to the task is presented with the goal of both simplifying the upkeep of the classifier as well as improving the performance. In the second topic the application of deep neural network classifiers in the context of a search for a new particle, the charged Higgs boson, is presented. Here the main focus is in producing a classifier that has been decorrelated from a variable of interest that will be used in making the final discovery or exclusion of the hypothetical particle. The classifier can then be used just like any other selection step in the analysis aiming to separate known Standard Model background events from the expected signal without distorting the distribution for the variable of interest. Both research topics present first time use cases at the CMS for deep neural networks in their respective contexts and the work done includes the full stack of solving a machine learning problem, starting from data collection strategy to cleaning the data and working out the meaningful input variables for the problem all the way to training, optimizing and deploying the model to get the final results for their performance.