Blair Bilodeau, graduate student at U of T’s Department of Statistical Sciences, received the Best Poster Award at the 14th Annual New York Academy of Sciences Machine Learning Symposium for his submission “Improved Bounds on Minimax Regret under Logarithmic Loss via Self-Concordance.”
The setting is probabilistic forecasting (e.g., assigning probabilities each day to whether it will rain the next day). Blair's work describes significantly improved estimates for minimax regret of sequential probability assignment under logarithmic loss against arbitrary expert classes (characterized by their metric entropy), and does so by giving the first argument in this setting to exploit the curvature of logarithmic loss to avoid an analysis that truncates the probability assignments away from 0 and 1.
This work was a joint collaboration with Daniel Roy, assistant professor at U of T, and Dylan J. Foster from MIT. The project was carried out during both the Simons Institute program on Foundations of Deep Learning last summer and the Special Year on Optimization, Statistics, and Deep Learning at the Institute for Advanced Study this winter.
The Department of Statistical Sciences at the University of Toronto continues to produce outstanding graduates who go on to have rewarding careers in Canada and across the world. We are proud of Blair’s outstanding work. Currently, he is a graduate researcher at the prestigious Vector Institute. He is expected to graduate from U of T with his PhD in Statistical Sciences in 2023. To learn more about his research, current work, and background, please visit Blair’s student profile on our website.