End-to-End Neural Networks for Top Quark Tagging Contain Hidden Representations of Physical Measurements

Artificial Neural Networks are increasingly used in high-energy physics analyses for picking signal events out from backgrounds and for trigger systems, among many other tasks. These networks are often treated as a "black box" where the behavior of the hidden layers is difficult to interpret for humans. Progress has been made in interpreting these layers using "linear probing," in which the output of each hidden layer is treated as an "intermediate representation" of the input data. These intermediate representations can be used to train linear regressors or "probes" which predict various quantities, revealing the kinds of models implicitly contained within these hidden layers. Using a Monte Carlo simulated sample of proton-proton collisions at 14 GeV, we train a Graph Neural Network to classify events as originating from either a top quark or the hadronization of a light quark or gluon, using only features that would be available in raw data such as energy of the final-state products, hence an End-to-End neural network. We show using linear probing that this network contains a linear representation of quantities such as the jet invariant mass of the event, despite never having seen these quantities in training. We also discuss potential applications of interpreting these layers in physics searches.