Accelerating Monte Carlo Bayesian Prediction via Approximating Predictive Uncertainty Over the Simplex