Statistical inference is a crucial step in gleaning insights from experimental / observational data to build better physics models to describe our universe. Whether it is to unravel the mysteries of what is happening in the interior of neutron stars, or to uncover the secrets about virtual Higgs bosons, precise inference requires a careful consideration of all sources of systemic uncertainties. Did we discover a new particle, or did we just discover a miscalibration of the detector? Such complex data analysis has traditionally involved multiple stages of lossy dimensionality reduction and simplifying assumptions in order to make the final inference feasible. This talk brings together Bayesians and frequentists in our quest for high-dimensional statistical inference to maximally leverage the collected data as well as prior physics knowledge to perform near-optimal and uncertainty-aware parameter inference. I will talk about a use case in Bayesian inference for neutron star astrophysics as well as a frequentist case for measuring the Higgs width, where we see a dramatic improvement in precision and interpretability with machine learning.
Bio
Dr. Aishik Ghosh is a researcher at UC Irvine and Berkeley Lab where he develops novel ML methods for particle and astrophysics. His current focus is on statistical inference, uncertainty quantification and debiasing, generative models and frugal AI. He obtained his PhD from the Université Paris-Saclay for developing simulation-based-inference methods and deploying the first deep generative model for fast simulation of the ATLAS detector. Aishik also works with the Organisation for Economic Co-operation and Development on matters of AI and science policy.