Learning Distributions for Physics: Sampling (FUND) and Inverse Problems (EUSSIP)
by
Prof.Vipul Arora(ESAT, KU Leuven)
→
Asia/Kolkata
A-304 and On Zoom
A-304 and On Zoom
Description
This talk is about sampling and inverse problems in computational Physics.
Deep generative models (such as normalising flows) complement Markov chain Monte Carlo methods for efficient sampling from high-dimensional distributions, but they require training samples to perform well. I will talk about our recently developed FUND algorithm, which reliably trains a normalising flow model without requiring any training samples. It obtains encouraging results for scalar phi^4 theory.
The second part of the talk is about simulation-based inverse problems. I will discuss our recent approach, EUSSIP, which employs uncertainty estimation along with active learning to iteratively search for the inverse solution. Experiments show improved performance for inverse Laplace transform and LAE tomography in cosmology.