Convergence Speed
ML-augmented reconstruction converges faster while reaching final Poisson NLL values comparable to the standard algorithm. Using iteration-to-epsilon, the ML method reduces wall-clock time by 2.1x to 2.3x. Early ML insertion at iteration 3…
1 sources - 4 claims
ML-augmented reconstruction converges faster while reaching final Poisson NLL values comparable to the standard algorithm. Using iteration-to-epsilon, the ML method reduces wall-clock time by 2.1x to 2.3x. Early ML insertion at iteration 3 or 5 gives the lowest convergence iterations in the hyperparameter study. The ML insertion can briefly increase Poisson NLL before gradient-based updates restore data consistency.