2025 workshop program
Leveraging Approximate Models for Exact Inference: A Hybrid AI-MCMC Approach
The work
Abstract
Modern cosmology and astronomy demand the modelling of systematics and quantities of interest with unprecedented precision and accuracy. The resulting data deluge necessitates innovative, scalable inference methods to efficiently explore the high-dimensional parameter spaces.
Although AI has significantly advanced astronomical and cosmological research, predictions from machine learning models while scalable often suffer from reduced accuracy and increased calibration challenges compared to their original counterparts due to their dependence on representative training data. I introduce novel methodology that accelerates Markov Chain Monte Carlo inference by utilizing and incorporating AI-approximated models. The core principle is a hybrid approach to exploit the computational efficiency of approximate AI models while ensuring the accuracy of the final results achieved using a combination with precise, non-emulated physical model predictions.