Michael Kearns

From SI410
Revision as of 11:42, 9 April 2021 by Tomea (Talk | contribs) (Issues of algorithmic fairness)

(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Michael Kearns is currently a professor and the National Center Chair in the Computer and Information Science Department at the University of Pennsylvania. He is the Founding Director of the Warren Center for Network and Data Sciences. He holds secondary appointments in the departments of Economics, Statistics of Wharton as well as their departments of Operations, Information, and Decisions. He is the founder and former director of Penn Engineering's Networked and Social Systems Engineering Program. Another former role of his being the co-director of Penn's interdisciplinary Institute for Research in Cognitive Science [1].

Michael Kearns.png



Dr. Kearns has prior work in the private sector. He has worked with the Bank of America, Lehman Brothers, SAC Capital, and Morgan Stanley in quantitative & algorithmic trading roles. He has experience working as an advisor to technology companies and venture capital firms. Kearns began a role with Amazon as part of their Scholars Program, focusing on algorithmic fairness, privacy, machine learning, and related topics within Amazon Web Services in June 2020[3].

In 2020, Kearns joined Amazon as a scholar due to his admiration for the company and his loyalty as an Amazon customer. Kearns' main goal working with Amazon is to implement the norms of the Amazon products and their services well between the Amazon Web Services (AWS) and the consumer side. The goal is to initiate internal and external conversations on the best way to implement those tasks between regulators, the policy makers, and other parties who are experts on hot topics for Amazon consumers[4].

Kearns is an elected Fellow of the American Academy of Arts and Sciences, the Association for Computing Machinery, the Association for the Advancement of Artificial Intelligence, and the Society for the Advancement of Economic Theory. [2]

Educational Background

Kearns's education began with his family. He was born into a highly academic family; his father, David R. Kearns, is Professor Emeritus at the University of California, San Diego in Chemistry (winner of the Guggenheim Fellowship in 1969), his uncle, Thomas R. Kearns, also a Professor Emeritus at Amherst College in Philosophy, his paternal grandfather Clyde W. Kearns was a well-known professor at the University of Illinois where his specialty was insecticide toxicology, and his maternal grandfather, Chen Shou-Yi, was a professor of History and Literature at Pomona College.

Michael Kearns completed his B.S. degree in math and computer science at the University of California at Berkeley graduating in 1985. In 1989, he completed a Ph.D. in Computer Science at Harvard University. His dissertation, "The Computational Complexity of Machine Learning," was published thereafter. Before joining AT&T Bell Labs in 1991, he continued with postdoctoral positions at the Laboratory for Computer Science at MIT and at the International Computer Science Institute (ICSI) in UC Berkeley. [2]


Kearns's research background is diverse, as he describes it in his University of Pennsylvania bio:

My research interests include topics in machine learning, algorithmic game theory and microeconomics, computational social science, and quantitative finance and algorithmic trading. I often examine problems in these areas using methods and models from theoretical computer science and related disciplines. While much of my work is mathematical in nature, I also often participate in empirical and experimental projects, including applications of machine learning to problems in algorithmic trading and quantitative finance, and human-subject experiments on strategic and economic interaction in social networks. [2] [5]

Research Articles

Kearns has published over 140 research papers dating back to 1987. Some of Kearns's most recent publications include "Differentially Private Query Release Through Adaptive Projection," ""Lexicographically Fair Learning: Algorithms and Generalization," and "Convergent Algorithms for (Relaxed) Minimax Fairness." These papers have been presented and published at the following conferences and journals: AAAI: Annual National Conference on Artificial Intelligence; AISTATS: International Conference on Artificial Intelligence and Statistics; ALT: Algorithmic Learning Theory; COLT: Annual Conference on Computational Learning Theory; EC: ACM Conference on Economics and Computation; FAT*: ACM Conference on Fairness, Accountability and Transparency; FATML: Fairness, Accountability and Transparency in Machine Learning; FOCS: IEEE Foundations of Computer Science; HCOMP: AAAI Conference on Human Computation and Crowdsourcing; ICML: International Conference on Machine Learning; IJCAI: International Joint Conference on Artificial Intelligence; ITCS: Innovations in Theoretical Computer Science; NIPS/NeurIPS: Neural Information Processing Systems; PNAS: Proceedings of the National Academy of Science; SODA: ACM Symposium on Discrete Algorithms; STOC: ACM Symposium on the Theory of Computation; UAI: Annual Conference on Uncertainty in Artificial Intelligence; WINE: Workshop on Internet and Network Economics.[6]


Michael Kearns and Aaron Roth holding "The Ethical Algorithm: The Science of Socially Aware Algorithm Design" [7]
Michael Kearns has written a total of 3 published books.

The Ethical Algorithm: The Science of Socially Aware Algorithm Design. The Ethical Algorithm is jointly authored with professor of Computer and Information Science at the University of Pennsylvania computer science department, Aaron Roth[8], and was published by Oxford University Press in 2019. Both Kearns and Roth are leading researchers in machine learning where their main emphasis of their research focuses in the real world application of algorithms and the overall design of these algorithms[9]. It is a general-audience book about the science of designing algorithms that embed social values like privacy and fairness, while still promoting the advance of data-driven scientific exploration.

An Introduction to Computational Learning Theory Jointly authored with Umesh Vazirani of U.C. Berkeley, this 1994 MIT Press publication is intended to be an intuitive but precise treatment of some interesting and fundamental topics in computational learning theory. The level is appropriate for graduate students and researchers in machine learning, artificial intelligence, neural networks, and theoretical computer science.

The Computational Complexity of Machine Learning The Computational Complexity of Machine Learning is published by the MIT Press. It is a revised version of Kearns' doctoral dissertation, which was completed in May 1989 at Harvard University, and was published as a part of the Association for Computing Machinery (ACM) Doctoral Dissertation Award Series. The publication details the scope of possibilities for efficient machine learning, in a mathematical study. [10]

Issues of algorithmic fairness

The book "The Ethical Algorithm" enlightens readers with many of Kearns' ideas on fairness, privacy, and equality with regard to the recent explosion of machine learning in our daily lives[11].

Issues and Notions of Algorithmic Privacy

While absolute data privacy is an ideal goal to some, with the idea that only aggregate data should be the only data shared, Kearns warns there are certainly costs to the complete privacy of data. The release of data is often what provides advancements in many of our daily technological applications. Industries like navigation apps, the medical field, and, broadly speaking, science all heavily rely on our data. If society decides to prohibit the release of all data, these core areas will have an exponentially harder time improving.

Kearns also rejects the concept of k-anonymity, an idea of anonymized data that was introduced by Latanya Sweeney and Pierangela Samarati and has recently gained traction among the computer scientist community. According to Kearns, it's another unnecessarily strict method of ensuring privacy.

Instead, Kearns offers up the idea of "differential privacy." This idea is that a person's data, in a specific context such as health data, is released up until it raises the probability of harm being done to them. So following the health example, a study should only make public data that has little impact on the probability of the study's subjects' health insurance going up.

Algorithmic Fairness

Kearns puts forth the idea that there are different types of fairness, and each may come in conflict with one another, creating an issue of deciding which notion should be the primary focus of optimization and an issue of balance between the three notions.

1.Statistical parity: The percentage of individuals who receive some treatment should be approximately equal across (our defined) groups. In simpler terms, this idea is that for an algorithm to be thought of as fair, each "group" of people impacted by the algorithm should have the same proportion of treatment. For example, an algorithm determining whether applicants have high enough credit for a home loan, should have the same percentage of approved applicants across race, gender, sexuality, etc. This can come in conflict with the next definitions of fairness.

2.Approximate equality of false negatives: The percentage of mistakes in terms of not giving treatment to individuals who might deserve it should be approximately equal across groups. Following the loan example, this notion would aim to have the same proportion of each race/gender/sexuality of applicants who are unjustly turned down in their application. This means some individuals may have a different decision on their application in the optimization of this definition compared to the first.

3.Approximate equality of false positives. Similar to (2), the percentage of mistakes in terms of giving treatment to individuals who did not deserve it (false positives) should be approximately equal across groups. Optimizing fairness in this sense will also bring about different loan decisions for some applicants than the optimization of the two prior definitions.

Scientific Reproducibility Issues

"The Ethical Algorithm" brings to light the "reproducibility crisis." This is the idea that many research findings within the data science community aren't being able to be verified by replicated studies. This is an issue that is usually caused by flawed methods within the research method that causes the researchers to publish false conclusions. Many aspects of research with data can be ambiguous and some researchers may choose to go about the process differently. When the researchers choose unscientific methods of choosing data or cleaning said data, for example, it can lead to false positives.

Kearns offers a method to clear up some of the ambiguity that comes with research data. He and the coauthor propose that data used by a researcher should be contained inside a database "operating according to a specific set of rules and researchers can only have access to results from through precise questions asked. Then the number of questions asked will also act as a correction for the hypothesis’ p-value. Such a method can deal with both scale and adaptability concerns." [12]

Morality and Scope of AI

As computers and algorithms hold more stake in our lives, Kearns warns we should be prepared in answering what amount are we comfortable with algorithms making decisions for us in our daily lives. Computers are increasingly spreading into important and complex aspects of daily life such as AI self-driving cars, automated warfare, and even banking decisions like the example from above of accepting/denying loan applications. Mistakes in any of these arenas are extremely impactful and we must determine if we are willing to risk the consequences of these potential mistakes against the convenience automation would provide.

With expansion of algorithmic automation into these subjects, Kearns is concerned with the problem of programming morality into these computers, which requires attempting coming up with answers to some questions of morality that humans themselves may not have cohesively agreed upon yet. Examples like: "The classic example is of a self-driving car faced with an inevitable fatal collision – Should the car take the action that minimizes harm to its owner / passenger or the one that minimizes total social harm?" [12]
  1. Dr. Michael Kearns https://www.turing.ac.uk/people/governance/michael-kearns
  2. 2.0 2.1 2.2 2.3 Bio: https://www.cis.upenn.edu/~mkearns/
  3. How I became a Quant: Insights From 25 of Wall Street's Elite http://edwardbetts.com/monograph/quantitative_trading_/_quantitative_%EF%AC%81nance/
  4. Michael Kearns Amazon Scholar, University of Pennsylvania https://www.amazon.science/scholars/michael-kearns/
  5. Michael Kearns External Professor: https://www.santafe.edu/people/profile/michael-kearns
  6. https://www.cis.upenn.edu/~mkearns/#publications
  7. EA pic: https://townhallseattle.org/event/michaels-kearns-and-aaron-roth/
  8. https://www.cis.upenn.edu/~aaroth/
  9. Amazon Scholars Michael Kearns and Aaron Roth discuss the ethics of machine learning https://www.amazon.science/latest-news/amazon-scholars-michael-kearns-and-aaron-roth-discuss-the-ethics-of-machine-learning/
  10. https://www.cis.upenn.edu/~mkearns/papers/thesis.pdf
  11. Michael Kearns Professor and national Center Chair https://www.cis.upenn.edu/~mkearns/
  12. 12.0 12.1 Ethical Algorithm: https://aristidouandreas.com/the-ethical-algorithm-aaron-roth-michael-kearns-review-and-summary/