The real world is nonstationary, the problem domain isn’t fixed and we can’t capture all variability in a static dataset or a learning environment e.g. our bodies change (& a robot’s body too), new words emerge in every language, etc.
How can we use causality and structure to enable knowledge composition in agents, allowing them to reason over explicit knowledge and make more interpretable decisions in a constantly changing world?
Discovering some, even imperfect, causal structure of the world can help agents reason, plan & explain (credit assignment) e.g. a person that knows how to drive on the right side of the road can drive on the left without re-learning how to drive, and largely benefit from systematic generalisation because human knowledge is compositional.
Reinforcement learning, Causality, Knowledge representation, Continual learning
Software Engineer @ Google
MSc Computer Science, ETH Zurich
BEng Computing, Imperial College London
Prof Özgür Şimşek
Prof Guy McCusker
