Gradient-based and perturbation-based methods can effectively quantify epistemic uncertainty in pre-trained neural networks without requiring the original training data or model modifications.
