evidential deep learning

Instead of sampling many times to estimate uncertainty, an evidential network outputs uncertainty directly in a single forward pass. It does this by predicting not a point answer but the parameters of a distribution over distributions: a Dirichlet over the class-probability simplex for classification, a Normal-Inverse-Gamma over the mean and variance for regression. The concentration of that higher-order distribution is read as how much evidence the model has.

The idea sits in the framework of subjective logic and evidence theory: a Dirichlet with large total concentration is a sharp opinion (high evidence, low epistemic uncertainty), while concentration near the uniform prior means I do not know. Sensoy and colleagues (classification) and Amini and colleagues (regression) train these with a loss that fits the data while a regularizer drives evidence toward zero on samples the model cannot explain, letting one network report both aleatoric and epistemic uncertainty cheaply.

The appeal is speed: no ensembling, no sampling, deterministic uncertainty at inference time. But the approach has drawn sharp criticism. Its epistemic-uncertainty signal is not derived from a proper posterior, can be poorly calibrated, depends strongly on the regularizer and the out-of-distribution training data, and may fail to grow on genuinely novel inputs. Treat it as a fast heuristic, not a Bayesian guarantee.