Uncertainty Quantification in Regression Neural Networks Using Likelihood-Based Belief Functions

We introduce a new method for quantifying prediction uncertainty in regression neural networks using evidential likelihood-based inference. The method is based on the Gaussian approximation of the likelihood function and the linearization of the network output with respect to the weights. Prediction uncertainty is described by a random fuzzy set inducing a predictive belief function. Preliminary experiments suggest that the approximations are very accurate and that the method allows for conservative uncertainty-aware predictions.