Monte Carlo simulation of physical systems with deep generative models



Shinichi Nakajima ( Technische Universität Berlin)

Shinichi Nakajima is a member of Machine Learning Group in Technische Universität Berlin. He received the master degree on physics in 1995 from Kobe university, and worked with Nikon Corporation until September 2014 on statistical analysis, image processing, and machine learning. He received the doctoral degree on computer science in 2006 from Tokyo Institute of Technology. His research interest is in theory and applications of machine learning, in particular, Bayesian learning theory, computer vision, and quantum chemistry.



Short Abstract:Deep learning has shown its usefulness in many fields of science. In this talk, we introduce our recent processes in Monte Carlo simulation in physics with deep generative models. Specifically, we extend our previous work, where we used generative models, e.g., autoregressive models and normalizing flows, that not only allow efficient sample generation but also provide the exact normalized sampling probability, and applied neural importance sampling for obtaining unbiased estimators. Our extension includes its application to estimating thermodynamic observables, such as free energy and entropy, in the lattice field theory, detection and mitigation of mode dropping, which can violates the condition for unbiasedness, and efficient training with path gradient estimators. Our experiments showed that our approaches can enhance the utility of deep generative models in physics, improve its reliability, and reduce the computational cost.