What is the universe made of? This is the core question particle physics aims to answer by studying the fundamental blocks of the universe. To study these blocks we require colliding particles at approximately the speed of light which produces high dimensional data in the order of peta-bytes per second, presenting considerable challenges in data processing and analysis. In order to validate or refute physical theories, it is necessary to distinguish the particles created in the collisions from the background noise. The data is processed through a complex pipeline with multiple non-interpretable data representations like images, sequences and graphs, at each level of processing. At the end of the pipeline there is a set of interpretable high-level features created using physics-motivated heuristics, which are analyzed using conventional statistical methods to make a classification. The multiple levels of data processing and representations opens the possibility of using techniques like deep learning to obtain improvements which in time will enable new discoveries in particle physics.

In this thesis, we show it is possible to bypass the dimensionality reduction step of traditional methods by using deep learning directly in the low-level detector data. This approach outperforms the-state-of-the-art methods in particle physics problems such as jet flavor classification, electron classification, and muon classification by 1.6%, 3.0% and 8.7% respectively. In addition, we show it is possible to achieve this performance using neural networks while maintaining the interpretability of high-level features, by using a recently developed technique to map the deep network into a space of physically interpretable high-level features that reproduce the performance of the deep network.