PHYSTAT Seminar: Extreme Lossless Data Compression for Likelihood-Free Inference

by Alan Heavens (Imperial)

Wednesday 10 Jan 2024, 16:00 → 17:30 Europe/Zurich

Likelihood-free inference (LFI), also known as simulation-based inference (SBI) is a technique that is growing in popularity in various fields, as the complexities of systematic errors, selection effects and so on often make a likelihood-based approach unfeasible.  The main challenge of LFI is the dimensionality of the problem.  The basic idea is to run simulations and keep only those that agree with the measured data, but if there are N>>1 data, then no simulations will come close, and LFI fails.  N needs to be reduced drastically, ideally to the number of model parameters, M, and if possible without any loss of information.  I will show how general analytic techniques of Extreme Data Compression such as MOPED, and neural-network-based compression can reduce N to M, and show that LFI can then do essentially loss-free inference from this typically very small set of numbers.  As an example, precise Type 1A supernova cosmology can be done with just 3 numbers.

Biography:  Alan Heavens is Professor of Astrostatistics at Imperial College, and was founding Director of the Imperial Centre for Inference and Cosmology.  With his collaborators, he has developed a number of novel statistical methods for analysing astrophysical data, including the use of the bispectrum to show that optical galaxies trace the matter distribution on large scales, optimised methods for non-gaussianity inference in the CMB, 3D weak lensing, extreme data compression, and large Bayesian Hierarchical Models for Cosmic Shear, with millions of latent parameters.

 

 

 

 

 

 

GMT20240110-150237_Recording_1790x828.mp4

PHYSTAT Seminar.pdf