DATA-EFFICIENT MACHINE LEARNING WITH MULTIPLE OUTPUT TYPES AND HIGH INPUT DIMENSIONS

Abstract

Recent research works in machine learning (ML) have focused on learning some target variables of interest to achieve competitive (or state-of-the-art) predictive performance in less time but without requiring large quantities of data, which is known as data-efficient ML. This thesis focuses on two highly related data-efficient ML approaches: active learning (AL) and Bayesian optimization (BO) which, instead of learning passively from a given set of data, need to select and gather the most informative observations for learning the target variables of interest accurately given some budget constraints. In particular, this thesis aims to (a) exploit the auxiliary types of outputs which correlate with the target variables for improving the learning performance of the target output type in both AL and BO algorithms and (b) scale up the state-of-the-art BO algorithm to high input dimensions.