Please note: This seminar will take place in DC 1304 and virtually over Zoom.
Reid
McIlroy-Young,
PhD
candidate
Department
of
Computer
Science,
University
of
Toronto
In the past half century Artificial Intelligence (AI) agents have grown to become embedded in modern life. Understanding why these AI agents make decisions has become increasingly difficult as they become more capable. These agents are components of larger sociotechnical systems making increasingly more significant decisions and are often doing so in collaboration with humans, for example, by providing evaluations to a judge or by making recommendations on the next movie to watch. As such, we need to build AI agents that are understandable to humans and that can understand humans with whom they interact.
In my talk I will present research done on chess. Chess has a long history of use at the forefront of AI research, from the mechanical Turk, to Deep Blue and AlphaZero. Moreover, chess allows access to large and diverse datasets, powerful chess engines that can act as oracles, and chess can be considered benign, i.e., the potential negative consequences of its research are low, making it an ideal model system for AI research. These characteristics led to my work on building a new style of AI chess engines, ones that attempt to mimic human play — instead of ones that achieve higher performance. These models, called Maia Chess, are able to predict the actions of a player at a given skill level with higher accuracy than any previous methods and have become the baseline for what it means to be human-like in chess. I will also present my work on the implications of these models towards bridging the gap between human and artificial intelligence, showing how they can be used to work as teachers finding mistakes, how they can be used to accurately identify people via “behavioural stylometry”, and the ethical considerations that arise from modeling human behaviour both in chess and in other domains.
Bio: Reid McIlroy-Young is a Ph.D. student in Computer Science at the University of Toronto, advised by Professor Ashton Anderson. His research explores how to make machine learning systems that can collaborate with humans and their implications for society. His best known work is Maia Chess, a project to use deep learning to simulate human-like play in chess.
Prior to the University of Toronto, he received a Master’s in Computational Social Science at the University of Chicago and bachelor’s degree in Mathematical Physics at the University of Waterloo.
To attend this seminar in person, please go to DC 1304. You can also attend virtually using Zoom at https://uwaterloo.zoom.us/j/96585505294.
For those attending virtually: The passcode will be provided by email on Friday before the seminar as well as on the morning of the seminar.