Abstract: Across many fields of science, computer simulators are used to describe complex
data generation processes. These simulators relate observations to the
parameters of an underlying theory or mechanistic model. In most cases, they are
specified as procedural implementations of forward stochastic processes
involving latent variables. Rarely do these simulators admit a tractable density
or likelihood function, thereby making inference difficult. The prevalence and
significance of this problem has motivated an active research effort in
so-called likelihood-free inference algorithms.
In this talk, we will discuss how likelihood-free inference has been carried for
decades in particle physics and then expand to new methods from machine
learning, including likelihood-ratio estimation through supervised learning [1]
and adversarial variational optimization [2].
Building upon these algorithms, we will then discuss how artificial intelligence
can enable the automation of the scientific method through the synergy of
three powerful techniques:
- Generic likelihood-free inference engines that enable statistical
inference on the parameters of a theory that are implicitly
defined by a simulator.
- Sequential design algorithms (e.g., Bayesian optimization) that balance
exploration and exploitation to efficiently optimize an expensive
black box function.
- Workflows that encapsulate scientific pipelines and extend the scope
from reproducibility to reusability.
[1] https://arxiv.org/abs/1506.02169
[2] https://arxiv.org/abs/1707.07113