Statistics of missed FRBs as a function of S/N from the CHIME Pathfinder. The histogram shows the distribution of 50,000 simulated FRBs in the test set (blue), as well as the events from that test set that were mislabeled as RFI by our frequency–time 2D CNN (orange). The false-negative rate goes to 0.5 for low S/N, as expected, since a binary classifier with no predictive power will classify correctly half of the time. The fraction of recovered events, or recall, gets close to 1 for high S/N.