Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence