Fig. 4: ROC curves showing the SMγγjj acceptance vs. H125 efficiency for different contaminated mixtures of H125 that is used to train the autoencoder.
From: Nanosecond anomaly detection with decision trees and real-time application to exotic Higgs decays
The legend indicates the percentage of the training sample consisting of H125 with the rest consisting of the SM sample, i.e., the uncontaminated is trained only on the SM sample. The 3 kHz line and the values for the uncontaminated autoencoder trigger and the 2γ trigger matches that of the top-right plot in Fig. 3. The plot is software-simulated results using bit integers as done in the firmware.
