USA - 2019
For foundational and breakthrough contributions to minimally-supervised learning.
For many real-world learning problems, unlabelled data is plentiful and easily harvested (images, video, clicks) but data with relevant labels for the task at hand (e.g., personalized learning, drug design) is much more expensive. Maria's foundational work developed the first theoretical framework that could capture the intuition behind the different types of learning methods designed to leverage both types of data and provided formal analysis in both qualitative and quantitative terms. At the same time, Maria also developed breakthrough results for active learning, where the learning algorithm intelligently chooses which data points to send for labelling. Most notably, she showed how active learning can always asymptotically improve over passive learning and gave the first general active learning algorithm that could tolerate realistic noise and model imperfections.