Andrey Malinin: Estimating Data and Knowledge Uncertainty

Data Fest Online 2020 Uncertainty Estimation in ML track Speaker: Andrey Malinin, Yandex & HSE In this video we will discuss how data and knowledge uncertainty can be estimate via ensembles of probabilistic models. We derive measures of data, knowledge and total uncertainty, showing that the average entropy of each model in an ensemble is a measure of data uncertainty, the entropy of the ensemble’ predictive posterior is a measure of total uncertainty, and that the ensemble’s diversity is a measure of knowledge uncertainty. Register and get access to the tracks: Join the community: