Modern Generative Methods for Bayesian Inference
Hosted by Astro Data Science Discussion Group
Downloads an .ics file · Times are in Europe/London · Google Calendar
Speaker: Gurjeet Jagwani, Institute of Astronomy, University of Cambridge
Series: Astro Data Science Discussion Group
Bayesian inference, increasingly in its simulation-based form, is the bread and butter for many workflows in astrophysics and cosmology. In parallel, modern generative methods like diffusion models and transformers have proven capable of solving problems of remarkable complexity. We will be walking through the inference of physical properties for around a hundred million galaxies from photometric surveys like the Kilo Degree Survey and Legacy Survey as an example, and show how modern generative methods can be utilised to address different components of Bayesian inference at this scale. We will also be talking about the limitations of these algorithms, specifically normalising flows and diffusion models. To address them, we have built two novel in-house algorithms: StAD and cStAD. Additionally, we will discuss how we have utilised agentic harnesses to aid our inference pipelines and future directions with "all in one" algorithms.
More like this near Cambridge
Going to Modern Generative Methods for Bayesian? Ask me anything about it.
I read the organiser's pages and answer in a few seconds.
Answers are AI-generated · Privacy