Speaker
Leande Thiele
(Kavli IPMU)
Description
Many inverse problems are only implicitly defined through a forward simulator.
In such cases, no closed-form likelihood function is known.
In cosmology, the canonical example is parameter inference from higher-order and field-level statistics.
By training neural networks on samples generated by the simulator, one can obtain an approximate likelihood and thus perform parameter inference.
I will present the main ideas in this simulation-based inference and put it in the context of current cosmological problems.
Then, I will turn to the problem of how to perform cosmological inference with realistic simulation budgets.
This issue necessitates multi-fidelity inference, which I will present some approaches to.
Primary author
Leande Thiele
(Kavli IPMU)
