26 April 2022
Maximilian Dax
#### Title:Real-Time Gravitational Wave Science with Neural Posterior Estimation Date: 26-04-2022 11:00-1200 Central European Summer time 
Speaker: **Maximilian Dax**, PhD Student at the Max Planck Institute for Intelligent Systems **Abstract:**
Inferring astrophyscial parameters from gravitational wave (GW) measurements is a central task in GW analysis. Standard inference methods, based e.g. on Markov chain Monte Carlo (MCMC), require days of computation for the analysis of a single GW event. In this talk I present our new approach DINGO that reduces this inference time to 20 seconds per event by using conditional normalizing flows. I then explain how physical symmetries can be used to enhance the accuracy of the inference networks. Finally, I demonstrate on real GW event data that our likelihood-free approach produces indistinguishable results from MCMC while being 1000 times faster. 1. Dax et al. Real-Time Gravitational Wave Science with Neural Posterior Estimation. Phys.Rev.Lett. 127, 241103 (2021) 2. Dax et al. Group equivariant neural posterior estimation. ICLR 2022 Watch Back › [1]: https://bereau.group/ [2]: /blog/ [9]: /contact/ [3]:https://github.com/undark-lab/swyft [4]:https://arxiv.org/abs/2011.13951 [5]:http://www.mathben.com/ [6]:https://pubs.acs.org/doi/10.1021/acs.jctc.0c00981 [7]:https://github.com/Ensing-Laboratory/FABULOUS [8]:www.evozyne.com
Speaker: **Maximilian Dax**, PhD Student at the Max Planck Institute for Intelligent Systems **Abstract:** Inferring astrophyscial parameters from gravitational wave (GW) measurements is a central task in GW analysis. Standard inference methods, based e.g. on Markov chain Monte Carlo (MCMC), require days of computation for the analysis of a single GW event. In this talk I present our new approach DINGO that reduces this inference time to 20 seconds per event by using conditional normalizing flows. I then explain how physical symmetries can be used to enhance the accuracy of the inference networks. Finally, I demonstrate on real GW event data that our likelihood-free approach produces indistinguishable results from MCMC while being 1000 times faster. 1. Dax et al. Real-Time Gravitational Wave Science with Neural Posterior Estimation. Phys.Rev.Lett. 127, 241103 (2021) 2. Dax et al. Group equivariant neural posterior estimation. ICLR 2022 Watch Back › [1]: https://bereau.group/ [2]: /blog/ [9]: /contact/ [3]:https://github.com/undark-lab/swyft [4]:https://arxiv.org/abs/2011.13951 [5]:http://www.mathben.com/ [6]:https://pubs.acs.org/doi/10.1021/acs.jctc.0c00981 [7]:https://github.com/Ensing-Laboratory/FABULOUS [8]:www.evozyne.com
COLLOQUIUM
colloquium