AF: Small: Foundations for Collaborative and Information-Limited Machine Learning

点击翻译按钮获取中文摘要

Machine learning increasingly is being used throughout society, and in a wide range of applications. Businesses use machine learning systems for decision support, internet sites use machine learning to better interact with users, our personal devices use machine learning to adapt to our needs, and our cars are beginning to use data-trained systems to improve safety. These applications bring up new opportunities as well as new concerns. Opportunities include the potential for systems to more rapidly learn and adapt through collaboration, and concerns include privacy and the fairness of algorithmically-made decisions. This project is aimed at developing new foundational understanding of these opportunities and concerns, to help guide the development of more efficient, more adaptive, and fairer, machine learning methods. This project additionally will support educational workshops on these issues, and more broadly will support the education and training of young scientists on these topics.Specifically, this project has the following four main thrusts: (1) Collaborative Machine Learning. How can devices with related learning tasks best collaborate to learn efficiently from only a modest amount of data, and how can privacy and related concerns be addressed? (2) Property Testing and Error Extrapolation. This thrust aims to develop methods that, from a small amount of labeled data, can reliably estimate how well a given learning algorithm or representation class would perform if given a much larger labeled data sample. (3) Semi-Supervised Learning. Semi-supervised learning refers to methods that combine labeled and unlabeled data, to learn well even when labeled data is limited. This work aims to develop theoretical foundations for an approach based on explicitly learning regularities within the unlabeled data and then using these to guide how learning is performed over the labeled data. (4) Fairness in Learning. There has recently been substantial concern about algorithmic decisions (such as whether to offer an applicant a loan) that could unfairly discriminate against certain classes of people. This work aims to develop improved theoretical understanding, tools, and guarantees for tackling these kinds of problems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.