Simulation-based inference, interpretability, and experimental design

Kyle Cranmer
New York University

I will consider experimental design from the point of view of statistical decision theory in a setting where a simulator provides a causal, generative model for the data. I will then review briefly the state of the art in simulation-based (or likelihood-free) inference techniques to highlight how computationally demanding this approach is and reveal what aspects of the scientific method are difficult to automate. The aim is that this will clarify which notions of interpretability are still important for science. I hope to both describe an optimistic path forward and to be realistic about our current capabilities.

Presentation (PDF File)

Back to Workshop II: Interpretable Learning in Physical Sciences