Performance metrics for the Sgr A^* and M87^* network training are displayed for various dedicated ZINGULARITY validation tests as described in Section 6. The validation error is computed from normalized labels of validation data not seen by the network during training. The mean absolute error (MAE) is computed as the average over all validation samples for normalized regression labels. The classification error (Class. error) is defined as one minus the network’s accuracy, i.e., the fraction of misclassified validation samples. For some studies, the classification errors get numerically close to zero beyond the logarithmic y-axis limit displayed in the figure. Training errors (not shown here) follow the validation curves with a constant (negative) offset, as is typical for neural networks (e.g., Advani & Saxe 2017). For the M87^* data, i_los and θ_PA are fixed. The training of the fiducial models is shown only up to the determined N_ep.