Properties of SuGOHI lenses in our test set that are recovered (plain) or missed (steps) by our networks, for fiducial score thresholds resulting in 50% recall. The histograms for all SuGOHI lenses are plotted in light gray. From left to right, we show the baseline ResNet, the baseline CNN, the ResNet trained on sets N1+L7 (network used in C21the ResNet trained on sets N1+L4 (natural θ_E distribution), the baseline CNN trained on griz-bands instead of gri, the CNN trained on sets N1+L2 (θ_E ≥ 1″), and the CNN trained on sets N1+L3 (θ_E < 2″).