Meta-learning causal reasoning

  • Dasgupta, Ishita
  • Invited Talk
  • [Slides]

Abstract

Causal reasoning is an integral part of human reasoning, and its absence in modern machine learning approaches has raised concerns about whether these approaches can lead to general artificial intelligence. In this talk, I will show how complex behaviors resembling causal reasoning arise via meta-reinforcement learning in unstructured deep learning architectures. Rather than incorporating explicit notions of formal causal reasoning, our approach demonstrates that causal reasoning can emerge as a reward-maximizing adaptation to a world containing causal structure. This gives rise to a more flexible notion of causal reasoning that can adapt to the specific domains in which an agent operates. This work lays the groundwork for causally-aware reinforcement learning agents, more efficient causal inference procedures, as well as new theories for how we understand causality in human cognition.

If videos are not appearing, disable ad-block!