Please read these instructions before posting any event on Fermilab Indico

Indico will be unavailable on Wed., Dec 18th from 7 - 7:30am due to server maintenance.

22–28 Jul 2018
Kellogg Hotel and Conference Center
EST timezone

Machine learning for multi-scale action matching

24 Jul 2018, 16:30
20m
103 (Kellogg Hotel and Conference Center)

103

Kellogg Hotel and Conference Center

219 S Harrison Rd, East Lansing, MI 48824
Algorithms and Machines Algorithms and Machines

Speaker

Prof. Phiala Shanahan (Massachusetts Institute of Technology)

Description

Critical slowing-down of HMC algorithms presents a significant challenge in achieving LQCD calculations at fine lattice spacings. A number of methods have been proposed that circumvent this issue by acting at multiple physical length scales, including perfect actions that aim to achieve almost-continuum physics at finite lattice spacings, and multi-scale thermalisation techniques. Such approaches require careful renormalisation-group matching of the LQCD actions defined at different scales such that they describe the same long-distance physics. An essential challenge is to solve the parametric regression task: Which action parameters best represent the coarse-scale physics of an ensemble of samples generated at a finer resolution, and vice-versa? Similar parameter regression problems of LQCD ensembles arise in the context of mixed-action LQCD simulations. I will discuss the applicability of machine learning to this regression task. Deep neural networks are found to provide an efficient solution even in cases where approaches such as principal component analysis fail. The high information content and complex symmetries inherent in lattice QCD datasets require custom neural network layers to be introduced and present opportunities for further development.

Primary author

Prof. Phiala Shanahan (Massachusetts Institute of Technology)

Presentation materials