ZINGULARITY performance diagnostics for various neural network training runs. 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 of 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. Panel a shows the fiducial models for Sgr A^* and M87^* in red and green, respectively. Here, we show the Sgr A^* model with a training length of 60 epochs; the 50 epoch model looks equivalent. For M87^*, validation errors are overall smaller compared to Sgr A^* and the classification errors get numerically close to zero beyond the logarithmic y-axis limit displayed in the figure here. Panel f also shows the training errors as gray curves.