Abstract

Deep learning is a popular tool for making inferences, as it has good performance in many practical applications. Since data scientists use deep learning, we should give theoretical guarantees on the quality of the constructed estimator. For some applications, it is not enough to have theoretical guarantees on the estimation error, but, in addition, practitioners need to quantify uncertainty. To quantify that, a practitioner can construct confidence sets. Researchers have only recently started giving theoretical guarantees on the accuracy of deep learning. The construction of confidence statements is still an open problem. In this talk, I will first go over the general ideas and concepts, discussing some earlier proposed methods for uncertainty quantification before diving into my contribution. I introduce a new Bayesian methodology: Empirical Bayesian deep neural networks (EBDNN). EBDNN is the first methodology with theoretical guarantees: the uncertainty quantification produced is valid from a frequentist point of view. Moreover, EBDNN is much faster to compute than alternative methods proposed for uncertainty quantification. Joint work with Botond Szabó.

Join Zoom Meeting https://maths-cam-ac-uk.zoom.us/j/99412853967?pwd=enFwNFpZcG0zM0o2WVlRQm1LdEttdz09

Meeting ID: 994 1285 3967 Passcode: 026816