CAREER: Towards a theory of machine learning with strategic interactions

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2). Machine learning (ML) algorithms use observed sampled data to uncover general patterns that can then be used for making predictions. Learning systems that interact with human data and stakeholders (such as those used in personalized medicine, content curation, financial markets, hiring, and lending) take place in a complex social and economic context. In this wide range of applications, there are feedback loops between learning algorithms and people that impact the quality of the learning process and the wellbeing of people. These feedback loops are currently not captured by the classical theory of ML, and handling them in an ad hoc, non-mathematical, way could have major social repercussions. This project will develop a rigorous mathematical framework for addressing interactions between learning systems and people and will draw from a wide range of academic traditions and fields, including Theory of Computing, Artificial Intelligence, Economics and Computation. This project also addresses educational and community building plans for enabling the next generations of students to contribute to a theory of machine learning for emerging and modern needs through cross-disciplinary research. This project also includes outreach, mentoring, and educational plans that will complement its technical goals, including a workshop series on "Learning in presence of Strategic Behavior" that brings together members of different communities and helps set an agenda for the field and "Learning Theory Alliance" that is a large-scale mentoring initiative for supporting the machine learning theory community. This project will build a theoretical foundation for ensuring both the performance of learning algorithms in the presence of everyday social and economic forces and the integrity of social and economic forces that are born out of the use of machine-learning systems. To achieve this, the investigator will consider adversarial, strategic, and collaborative interactions. For adversarial and long-term strategic interactions, the project will explore online decision processes and contribute online learning algorithms that perform well in presence of more realistic adaptive and non-myopic strategic agents. The project also explores the long-term social impact of strategic play and communication on learning and quality of available information, with an eye towards understanding and addressing biased and polarized beliefs. Additionally, to reap the full benefit of collaborative interactions, the project will align the performance of learning algorithms with the needs and preferences of participating agents. This will lead to the design of collaborative learning protocols that are differentially private and statistically efficient. 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.