Speaker
Description
Modern jet taggers operating directly on jet constituents have shown impressive performance, yet the high-level physics features they exploit to outperform classical methods remain only partially understood. In this paper, we discuss hidden particle flavor information imprinted in the kinematics of jet constituents; this flavor information is less accounted for in constructing high-level features that rely solely on the constituent kinematics. We argue that the flavor information and constituent kinematics are not independent because of angular resolution heterogeneity and particle-dependent response of detector subsystems, and hence, the modern jet taggers can exploit this flavor information even when the constituent-type information is masked out. To demonstrate accessibility of the flavor information, we first compare two Particle Transformers classifying flavor-specific top jets in a constituent type-blinded setup, i.e., $t \rightarrow bu\bar{d}$ jets vs.~QCD jets and $t \rightarrow bc\bar{s}$ jets vs.~QCD jets. The flavor asymmetry in the classification performance is a sign that Particle Transformer has access to the hidden flavor information. As the particle flow object type is the key detector-level quantity that carries the flavor information, we then explicitly show that object types can be inferred from the kinematics of neighboring constituents in order to demonstrate a correlation between flavor and kinematics. Our findings provide insight into why modern jet tagging architectures perform well, and highlight the importance of understanding the interplay between particles and detector effects when interpreting highly capable neural networks for jet classification.