Description
A method is presented to extract salient information from a deep neural network classifier of jet substructure tagging techniques, using expert variables that augment the inputs, using layerwise relevance propagation. The results show that these eXpert AUGmented (XAUG) variables can be used to easily interpret the behavior of the classifier, and in some cases can capture the behavior of the classifier completely. This can be used both to understand the behavior of complicated classifiers, and also to utilize them to guide development of expert variables that can encapsulate the physics the classifier is learning.
Authors
Christine Angela McLean
(SUNY Buffalo)
Salvatore Rappoccio
(The State University of New York SUNY (US))
Lauren Meryl Hay
(SUNY Buffalo)
Garvita Agarwal
(The State University of New York SUNY (US))
Margaret Morris
(University at Buffalo)
Ia Iashvili
(The State University of New York SUNY (US))
Mannix Benjamin
(University at Buffalo)
Ulrich Schubert
