Abstract: 

Humans have a profound ability to generalize from a small number of noisy examples, and yet, our best AI methods typically require large amounts of training examples and time to achieve satisfactory performance. This challenge is exacerbated in many important real-world domains where the data is complex, and thus, converging to a good solution is difficult even when large sample sets are available. In my work, I am inspired to solve challenging and important problems connected to social good, such as health and safety, medicine, the environment and green materials, and have utilized these domains to advance the understanding of the potential and limitations of current AI. In particular, my research focuses on the development and understanding of algorithms for domains in which samples are limited and / or costly. In this talk, I will discuss how small sample sets paired with challenging data properties impact the predictive performance of machine learning classifiers and knowledge discovery. I will present my previous work on handling extreme imbalance for classifier learning. In addition. I will discuss my recent work on a reinforcement learning paradigm for scientific discovery. The framework utilizes model uncertainty to improve sample efficiency and empowers the agent to proactively decide when to act and when to sample the state.

Bio:

Colin Bellinger is a Research Scientist in Data Science for Complex Systems at the National Research Council of Canada. He received a Ph.D. from the University of Ottawa in 2016. Dr. Bellinger held a Postdoctoral Fellowship with the Alberta Machine Intelligence Institute (AMII) at the University of Alberta in Edmonton, Canada, and the Donald Hill Fellowship in Computer Science at Dalhousie University in Halifax, Canada. He has co-authored more than 20 articles in international conferences and journals focused on machine learning, data mining, health and computational security, and received two best paper awards. Dr. Bellinger was the Program Chair for the Graduate Students’ AI Symposium in 2017 and 2019, has participated on numerous program committees and reviewed for over 10 international journals.