David Scott Krueger

My research is focused on deep learning, AI safety and alignment, and more specifically, with understanding how to communicate or “specify” what behavior is desired.  I’m pursuing the three research directions I view as most promising in this area:

• Learning what humans want from human feedback (e.g. via reward modelling).
• Managing the incentives an AI system has to influence the world (e.g. to prevent user manipulation in content recommendation systems).
• Getting deep nets to understand the world the same way people do (e.g. so that they can solve out-of-distribution generalization problems).

I believe solving these problems will ultimately be crucial to our ability to harness advances in AI to improve the world.  Ultimately, however, whether AI will be developed and used beneficially depends more on social than technical factors.  I’ve written about this for the Partnership on AI, and in a recent report on trustworthy AI development.  At Mila, I’ve run reading groups on AI Ethics and mechanisms for tackling global governance issues.

Some highlights of my previous research include:

• A Closer Look at Memorization in Deep Networks,
• Developing two RNN regularization methods,
• Developing methods for Bayesian deep learning using normalizing flows.