image/svg+xml Safe area for synthetic points. Let us call it S(x) = { x + ζ : ∥ζ∥ ≤ ε } / {y + ζ : y ∈ Dmaj , ∥ζ∥ ≤ δ} unsafe area for synthetic points D = Dmaj ∪ Dmin with Dmin = {x1, x2, …} Questions: Can we train a generator so that it produces points respecting S(x) for a given x? The generator learns the shape of the safe area. noise → x → →v with x + v ∈ S(x) G