Explainable AI for ML jet taggers using expert variables and layerwise relevance propagation

A bstract framework is presented to extract and understand decision-making information from a deep neural network (DNN) classifier of jet substructure tagging techniques. The general method studied provide expert variables that augment inputs (“eXpert AUGmented” variables, or XAUG variables), then apply layerwise relevance propagation (LRP) networks both with without variables. are concatenated the intermediate layers after network-specific operations (such as convolution recurrence), used in final network. results comparing addition show can be interpret behavior, increase discrimination ability when combined low-level features, some cases capture behavior completely. LRP technique find relevant using, rank allowing one reduced set features part performance. In studies presented, adding DNNs increased efficiency classifiers by much 30-40%. performance improvements, an approach quantify numerical uncertainties training these presented.