PI3NN: Prediction intervals from three independently trained neural networks

We propose a novel prediction interval method to learn prediction mean values, lower and upper bounds of prediction intervals from three independently trained neural networks only using the standard mean squared error (MSE) loss, for uncertainty quantification in regression tasks. Our method requires no distributional assumption on data, does not introduce unusual hyperparameters to either the neural network models or the loss function. Moreover, our method can effectively identify out-of-distribution samples and reasonably quantify their uncertainty. Numerical experiments on benchmark regression problems show that our method outperforms the state-of-the-art methods with respect to predictive uncertainty quality, robustness, and identification of out-of-distribution samples.