Amazon Scholar, University of Pennsylvania
Michael Kearns, a professor in the Computer and Information Science department at the University of Pennsylvania, where he holds the National Center Chair, joined the company as an Amazon Scholar in June 2020. Kearns joined the Penn faculty in 2002. He also holds secondary appointments in the Economics department, as well as the departments of Statistics, and Operations, Information and Decisions (OID) within Penn’s Wharton School of Business.
Kearns is the founding director of the Warren Center for Network and Data Sciences, and the faculty founder and former director of Penn Engineering’s Networked and Social Systems Engineering program. He’s also a faculty affiliate in Penn’s Applied Math and Computational Science graduate program, and until July 2006 was the co-director of Penn’s interdisciplinary Institute for Research in Cognitive Science. In August 2018, Kearns became an external faculty member at the Santa Fe Institute.
Kearns is a leading researcher in machine learning, algorithmic game theory, algorithmic trading and related topics. He also has interests in computational social science, and differential privacy.
Here, Kearns describes some of his previous research, and what he’ll be focusing on as one of the newest members the Amazon Scholars program.
Q. In 1994, you and colleague Umesh Vazirani published An Introduction to Computational Learning Theory, which has become a standard text for the mathematical study of the design and analysis of machine learning algorithms. In that book you discuss the Probably Approximately Correct (PAC) model of learning. Can you explain the main thesis of PAC?
A. The PAC model was the first mathematical framework to put machine learning on firm algorithmic foundations. It allows comparison of different algorithms with respect to resources, like computation time, and sample size required. I think one of the reasons for its great success was the integration of the statistical aspects of learning with the algorithmic aspects. In addition to elucidating some of the barriers to efficient learning, it also spawned some very practical methods, such as Boosting.
Q. Twenty-five years later, you and colleague Aaron Roth authored The Ethical Algorithm: The Science of Socially Aware Algorithm Design. In the book, you make the point that understanding and improving the algorithms that run our lives is rapidly becoming one of the most pressing issues of this century, and you offer a new approach to the science of socially aware algorithm design. Can you provide an overview of the new approach Aaron and you are advocating?
A. Yes, our book is about the science of designing algorithms that literally embed social norms such as fairness and privacy into their code. Of course, the first step in this process is being mathematically precise about what we mean by such norms, and this is often challenging, but also revealing.
Once we’ve chosen a definition of fairness, for example, we can design our algorithms to obey it. In a consumer lending application, we might design a machine learning algorithm that doesn’t simply minimize predictive error, but minimizes predictive error subject to a fairness constraint, like equalizing the false rejection rates across different racial groups. This constraint will have costs --- in particular, higher predictive error --- but such tradeoffs between accuracy and fairness are unavoidable. In a similar vein, differential privacy is an important technology for designing algorithms providing privacy guarantees to individual citizens’ data, and managing the inevitable tradeoffs between algorithmic accuracy and privacy.
We generally argue that socially aware algorithm design is an important and necessary complement to legal and regulatory efforts to constrain the negative consequences of machine learning.
Q. As an Amazon Scholar, you will be working on a few initiatives related to machine learning fairness and privacy. Can you provide an overview of the work you’ll be directing?
A. First of all, I’m very excited to be joining Amazon; as a longtime customer and admirer, I look forward to collaborating with all the amazing scientists, engineers and business leaders here. I’m hoping to play both leading and supporting roles in Amazon’s efforts in algorithmic fairness and privacy across the spectrum --- from the implementation of such norms in Amazon products and services on both the AWS and consumer side, to internal and external discussions of how best to do so, to interactions with regulators, policy makers and others outside the company on these important topics for society.