- How to design efficient optimisation methods. Such methods would not only lead to fast convergence (in terms of examples), they would also be amenable to distribution. See for instance SAG.
- What are principled ways of using the output of related tasks to extract information from weakly labeled examples.Using this data could drastically reduce the amount of labeled data required. See for instance this paper.
- How to make best use of past experience in reinforcement learning. Off-policy techniques are more efficient provided the variance can be controlled, an issue similar to that of stochastic gradient methods. See for instance this paper.
In the past, I've also worked on ML systems for auctions and am generally interested in mechanism design for auction systems.
Since August 2017, I am also an adjunct professor at McGill.