Mouhssine Rifaki

Research

I work on reinforcement learning that does not need millions of trials. Today's algorithms need that many because they have to rediscover the structure of every problem from scratch. My three current angles are: extracting that structure more directly, doing it in multi-agent settings where it is hardest to see, and building theory for how fast the cost of learning falls. The shared hypothesis is that the right inductive biases on data and architecture should reduce sample complexity by orders of magnitude rather than constant factors. Alongside that, I am also interested in deep learning theory.