LSSTDESC / SprintWeek2020

Meeting repository for the LSST DESC 2020 Sprint Week
Apache License 2.0
2 stars 0 forks source link

[SPRINT] Bayesian Variational Autoencoder cosmological parameter inference pipeline #20

Open mdschneider opened 3 years ago

mdschneider commented 3 years ago

Bayesian Variational Autoencoder cosmological parameter inference pipeline

Build a deep learning based emulator of cosmological parameter estimation MCMC to support new types of LSS summary statistics (such as lensing mass maps).

Contacts: Michael Schneider, Ryan Dana, Bob Armstrong @rearmstr Day/Time: Tues-Fri PST sprint hours Main communication channel: desc-wl-pipelines GitHub repo: https://github.com/mdschneider/cosmo_bcvae Pitch slides: https://confluence.slac.stanford.edu/download/attachments/279347747/DESC_sprint_BCVAE_20201201.pdf?api=v2

Goals and deliverable

Demonstrate that we can obtain comparable posteriors in Omega_m and sigma_8 using both MCMC and the new BCVAE pipeline given simulated tomographic gal-gal or shear-shear two-point functions.

Resources and skills needed

We have an initial BCVAE implementation to get started and are working to compare with existing examples in the Cobaya code.

Skills needed:

Detailed description

Eventually, we aim to build and demonstrate cosmological parameter inference pipelines for non-traditional summary statistics of large-scale structure that may contain more information than the 3x2 (galaxy and shear) two-point functions. We have previously proposed to use weak lensing mass maps directly in a Bayesian forward-model to constrain cosmological parameters by simulating full LSS light-cones within a Markov Chain Monte Carlo analysis. Such an approach can quickly become computationally intractable (although there are some successes in the literature). We aim to explore recently published advances in auto-encoders to use deep learning to produce accurate multivariate posterior distributions.