I'm a PhD student at Mila and University of Montreal, advised by Aishwarya Agrawal. This summer I'll join Qualcomm Research in Amsterdam as an intern. I work on generative models, world models, and reinforcement learning.

Research

Scaling on internet text has taken us remarkably far — further than most of us expected. But the next leap, the one that gets us to general and physical intelligence, will come from agents that learn by acting in the world. This is what the wisest minds in the field are starting to call the era of experience: the data that matters is no longer what humans wrote down, it is what an agent generates for itself by trying things. To get there, we need models that can simulate the world well enough to plan, fail, and try again inside it.

I think the path runs through the generative paradigm. The models we already use to generate images and video are, with the right training signal, world models. To get them to work for an agent in a big world, we need better representations and better policy gradient algorithms for scaling experience. That is what I work on.

A few deep-learning "recipes" I keep coming back to: the target network trick from RL, which later echoed through SSL and flow-based models; and asymmetric views in self-supervised learning.

Selected publications

See Google Scholar for the full list. *denotes equal contribution.

Misc

Talks

Slides from a few talks I put extra care into. Mostly for our group at Mila.

Academic Service