Ability of a model trained on simulated bursts to correctly identify Galactic pulsars. Our DNN classifier’s output is a list of triggers ranked by their probability of being an FRB. Shown are frequency (ordinate) vs. time (abscissa) arrays of the test triggers, identical to the top panels in Figure 5. The set includes real pulsars. Events that the classifier thinks are FRBs (p > 0.5) are boxed in black, and non-FRBs are boxed in red. The most likely events are in the top two rows; the marginal events are in the middle row, where the predicted labels transition from “FRB” to “RFI”; and the bottom row shows the events least likely to be true positives. In this case, we trained on 4850 known false positives from the CHIME Pathfinder and 4850 simulated FRBs. The test data are a separate set of several hundred triggers consisting of known false positives, single pulses from B0329+54, and giant pulses from the Crab. The Pathfinder classifier gets fewer than 1% wrong. Our classifier trained on Apertif data achieves 99.7% recall.