Evolution of the network output prediction when modeling an example localized perturbation to a smooth lens potential. Mock noisy lensing image data and the true perturbation are shown on the left. The predicted lensing image, the network reconstruction of ψ_pert, and residuals between the data and the predicted image at different training epochs are also shown. Training is performed via back-propagation of the gradients through the simulator, as presented in Fig. 2.