Comparison of the performance of the three networks described in Sect. 4. All samples include doubles and quads and a weighting factor of w_θE = 5, but different Einstein radius distributions or lower limits on the Einstein radius (see legend). Bottom row: normalized histograms of the difference between predicted values and ground truth for the five parameters and above the 2D correlations distribution: 1σ contour (solid line) and 2σ contour (dotted line).