Can a single neuron learn predictive uncertainty?
Published in Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 2021
Wait, wait, the Fermi-Dirac distribution function is what ML people call sigmoid? Isn’t this what is sometimes used as activation of artificial neural units? I was (and still feel) new to this community. However, following this train of thought, a novel method for estimating quantiles of a distrubution and then prediction intervals for regression problems was born.
The method is pretty cheap and accurate. It can easily be extended for the estimation of confidence intervals for the accuracy (and other metrics) in classification problems.