Grant: Discovery Grants Program - Individual

Machine learning is playing an increasingly central role in decision making systems in fields as diverse as autonomous driving, robotics, commerce, healthcare and medicine. These successes, however, have largely relied on human supervision in the form of accurate annotation of sufficient training data, which incurs significant cost, and is often the primary barrier to practical deployments. Additionally, machine learning must increasingly address concerns about data privacy and security while coping with ever more decentralized data storage and processing systems. Therefore, this research will investigate deeper data exploitation strategies that enable prediction model training in more realistic annotation availability scenarios. One key direction is to extend domain adaptation methods to preserve privacy and cope with distributed learning. Domain adaptation seeks to reduce the annotation cost of learning a prediction model in a target domain by transferring information from an annotation-rich source domain. However, accessing data from both the source and target domains violates privacy constraints. Therefore, this research will develop new domain adaptation methods that can perform knowledge transfer without accessing the source data, while also coping with federated learning scenarios where data is stored in a decentralized manner. A second challenge arises from the fact that auxiliary annotation-rich sources are typically not available in high expertise environments, such as medical data analysis. Such scenarios raise a bold new challenge, exemplar learning, where only a few richly annotated examples are available for learning a prediction model. To address this challenge, new data synthesis and augmentation methods will be developed that can enrich the limited annotation, and new self-training methods will be designed for leveraging unlabelled auxiliary data. This research will fundamentally reduce annotation cost in specialized image analysis problems. Finally, although accurate data annotation is expensive to obtain, low-cost weakly annotated data can often be collected through crowdsourcing or weak supervision mechanisms in some domains. This research will therefore develop new weakly supervised learning methods that can exploit low-quality imprecise annotations in prevalent learning scenarios, such as semi-supervised learning, domain adaptation, and federated learning. Overall, this research program will provide new data exploitation techniques that expand the applicability of machine learning to more practical annotation availability scenarios. This research will dramatically reduce the dependence of supervised machine learning on accurate manual annotation, broaden the availability of usable supervision resources, and alleviate the barrier created by domain expertise. By addressing both application-driven and methodology-driven aspects, this program will contribute advances to both applied and fundamental machine learning.