Séminaire Univers / |
| « Observational Inference with Machine Learning: Investigations in Galaxy Cluster Mass Estimation » |
|
Matt Ho |
| Recently, Machine Learning (ML) approaches have shown improvement upon traditional analytical methods for observational inference problems. These gains are driven by learning to model high order data features directly from cosmological simulations. However, these data-driven methods are conditional on assumptions used to generate training data and their naive application can result in predictive biases. In this talk, I will discuss aspects of ML solutions to observational inference problems and steps for building robust validation pipelines to ensure reliable predictions. As a case study, I will describe our previous work in leveraging the use of machine learning models to infer dynamical cluster masses from spectroscopic surveys. Using Convolutional Neural Networks (CNNs), we built models which substantially mitigated systematics in the virial scaling relation and produced dynamical mass estimates of galaxy clusters with remarkably low bias and scatter. I will detail our development process, from inference to uncertainty quantification to verification on real systems such as the Coma and CLASH clusters. Lastly, I will describe how these learned lessons will guide our current work on large scale Simulation Based Inference of galaxy surveys done within the Learning the Universe collaboration. |
|
mardi 15 novembre 2022 - 11:00 Salle des séminaires Évry Schatzman Institut d'Astrophysique de Paris |
| Pages web du séminaire / Seminar's webpage |