Classification of jets as originating from light-flavor or heavy-flavorquarks is an important task for inferring the nature of particles produced inhigh-energy collisions. The large and variable dimensionality of the dataprovided by the tracking detectors makes this task difficult. The currentstate-of-the-art tools require expert data-reduction to convert the data into afixed low-dimensional form that can be effectively managed by shallowclassifiers. We study the application of deep networks to this task, attemptingclassification at several levels of data, starting from a raw list of tracks.We find that the highest-level lowest-dimensionality expert informationsacrifices information needed for classification, that the performance ofcurrent state-of-the-art taggers can be matched or slightly exceeded bydeep-network-based taggers using only track and vertex information, thatclassification using only lowest-level highest-dimensionality trackinginformation remains a difficult task for deep networks, and that addinglower-level track and vertex information to the classifiers provides asignificant boost in performance compared to the state-of-the-art.
展开▼
