I am an Associate Professor of Computer Science at the University of Toronto, and a founding member of the Vector Institute. I am also a Member of Technical Staff on the Alignment Team at Anthropic through the summer of 2024.

My research has focused on better understanding neural net training dynamics, and using this understanding to improve training speed, generalization, uncertainty estimation, and automatic hyperparameter tuning. I’m now focusing on applying our understanding of deep learning to AI alignment. Given how fast AI is progressing, the problem of ensuring AIs are robustly aligned with human values seems like the most important thing we can be working on now. Specific directions I’m interested in include:

• Given a surprising behavior from an AI system, how can we efficiently figure out which training examples were responsible (so that we can then investigate the detailed mechanisms)?
• When do neural nets learn mesa-optimizers, i.e. when is the forward pass doing some sort of optimization without us building it in? Do such cases lead to different generalization patterns? How can we detect this, reverse engineer the inner objective function, and/or prevent it from happening?
• How can we elicit reliable information from models we don’t fully trust?

I hold a Sloan Fellowship, Canada Research Chair, and Canada CIFAR AI Chair. Previously, I received my BS in symbolic systems from Stanford in 2008, my MS in computer science from Stanford in 2009, and my PhD in computer science from MIT in 2014, studying under Bill Freeman and Josh Tenenbaum. From 2014-2016, I was a postdoc at the University of Toronto, working with Ruslan Salakhutdinov. Along with Colorado Reed, I created Metacademy, a web site which uses a dependency graph of concepts to create personalized learning plans for machine learning and related fields.

Contact

Department of Computer Science
University of Toronto
Office: Pratt 290F
6 King’s College Rd.
Toronto, ON M5S 3G4, Canada
Phone: 416-978-7391
e-mail: rgrosse_at_cs_toronto_edu