Séminaire Doctoral / Seminar PhD |
| « Lightening black-box models » |
|
Tristan Hoellinger |
|
Next-generation large-scale optical surveys, such as Euclid and LSST, promise to unveil the nature of dark energy, elucidate the processes driving cosmic inflation, and constrain neutrino masses with unprecedented precision. Despite the wealth of data these surveys will provide, they will be dominated by systematic rather than statistical uncertainty, challenging standard Bayesian inference methods that heavily rely on the accuracy of the forward models involved. Notably, even slightly misspecified models can significantly bias or overconcentrate the credible posteriors on cosmological parameters, risking false discoveries. Furthermore, even if systematic effects arising from survey strategies were fully understood and controlled, theoretical challenges like galaxy biasing and non-linear structure growth at late times would persist, reducing the detectability of cosmological signatures and therefore calling for novel inference techniques.
In this talk, I will present recent methodological advances to perform implicit likelihood inference of cosmology using arbitrarily complex black-box models of cosmological surveys. These advances enable efficient and thorough checks for systematic effects by leveraging our theoretical understanding of the primordial matter power spectrum after inflation, or any other latent variable in the forward model of cosmological observables. |
|
vendredi 7 juin 2024 - 16:00 Salle des séminaires Évry Schatzman, Institut d'Astrophysique |
| Page web du séminaire / Seminar's webpage |