Speaker
Sophia Vent
Description
Modern ML-based taggers have set new benchmarks for jet classification tasks at the LHC, surpassing traditional algorithms in performance. However, their opaque decision-making processes pose challenges for interpretability. In this work, we investigate what a low-level tagger learns when trained on quark-gluon discrimination. We identify a small set of learned latent features that correlate strongly with physics observables. Remarkably, only three latent features are sufficient to capture the full discriminative power. Moreover, we apply symbolic regression to derive compact analytic expressions that approximate the tagger output in terms of interpretable features.