Michael Kearns is a professor at University of Pennsylvania and a co-author of the new book Ethical Algorithm that is the focus of much of our conversation, including algorithmic fairness, bias, privacy, and ethics in general. But, that is just one of many fields that Michael is a world-class researcher in, some of which we touch on quickly including learning theory or theoretical foundations of machine learning, game theory, algorithmic trading, quantitative finance, computational social science, and more. This conversation is part of the Artificial Intelligence podcast. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter, LinkedIn, Facebook, Medium, or YouTube where you can watch the video versions of these conversations. If you enjoy the podcast, please rate it 5 stars on Apple Podcasts or support it on Patreon. This episode is sponsored by Pessimists Archive podcast. Here's the outline with timestamps for this episode (on some players you can click on the timestamp to jump to that point in the episode): 00:00 - Introduction 02:45 - Influence from literature and journalism 07:39 - Are most people good? 13:05 - Ethical algorithm 24:28 - Algorithmic fairness of groups vs individuals 33:36 - Fairness tradeoffs 46:29 - Facebook, social networks, and algorithmic ethics 58:04 - Machine learning 58:05 - Machine learning 59:19 - Algorithm that determines what is fair 1:01:25 - Computer scientists should think about ethics 1:05:59 - Algorithmic privacy 1:11:50 - Differential privacy 1:19:10 - Privacy by misinformation 1:22:31 - Privacy of data in society 1:27:49 - Game theory 1:29:40 - Nash equilibrium 1:30:35 - Machine learning and game theory 1:34:52 - Mutual assured destruction 1:36:56 - Algorithmic trading 1:44:09 - Pivotal moment in graduate school
Michael Kearns is a world-class researcher in many fields, including learning theory, game theory, quantitative finance and computational social science. This podcast looks at episodes in history where something new was introduced and explores why it freaked everyone out, with the latest episode on mirrors and vanity in the modern day of the Twitter world.
The prevalence of abuse of power can be avoided through awareness of different social norms in various worlds, rather than assuming that only evil people rise to power. The implementation of algorithmic solutions can differ between computer scientists and philosophers in their perspective on fairness.
The speaker discusses his technical definition of fairness in finance and how it relates to variances in market positions and the denial of loans, as well as ongoing human subject experiments exploring fairness in this space.
The episode discusses fairness guarantees for demographic properties such as race, gender, and age in AI systems, especially in the context of criminal sentencing and parole decisions, which remain unresolved issues in the debate over affirmative action.
Predictive models that are trained to maximize overall accuracy will be less accurate among minority groups if the training data set is mainly composed of white males, pointing out the trade-offs between fairness and accuracy that often exist in these models.
The podcast discusses the negative impact personalization through machine learning has had on society by creating filter bubbles and echo chambers, and suggests tuning the algorithm to show opposing viewpoints.
The speaker is leery of letting algorithms explore alternative definitions of fairness, and believes that social media platforms are currently in a bad equilibrium. They believe that while machine learning has had great successes, these successes have been on a narrow set of tasks.
While data anonymization is a widely adopted technical solution for data privacy, it has several fundamental flaws that make it susceptible to re-identifying people, especially when combined with other anonymized data sets and publicly available information.
Differential privacy is a method used in data analysis that adds noise to the actual value output to prevent reverse engineering of the data. This ensures privacy without compromising on the accuracy of the results obtained from statistical modeling and machine learning.
The speaker is optimistic about finding a better compromise in balancing individual data privacy and commercial uses of data. There is a need for individual privacy guarantees and control of their data despite the weak guarantees and little control at present.
In this episode, the guest explains his work in the field of algorithmic game theory, looking at settings in which the number of actors is potentially large, and still needing algorithmic ways of predicting or influencing what will happen in the design of platforms, such as navigation apps.
This podcast explores the different types of high-frequency trading including statistical arbitrage and optimized execution, their differences and how they're related. It also highlights the importance of low level granular buying and selling data in the exchange.
The process of transitioning from being a student of what's been done before to doing your own thing and figuring out your interests and strengths as a researcher is a common one for doctoral students. This requires carving out a career with independence and choosing what you like to think about.