Speaker
Denis Boyda
(ANL, MIT)
Description
Flow-based models were previously successfully applied for the generation of configurations in lattice (gauge) field theories in two dimensions. In this work, we discuss further development of this approach for lattice gauge theories in four dimensions. We show several implementations and apply improvements to the approach. We study different masking patterns and choices of frozen loops, as well as improvements to the flow transformations. Also, we investigate different machine learning methods such as different architectures of neural networks and training protocols.
Authors
Dr
Daniel Hackett
(MIT)
Danilo Jimenez Rezende
(DeepMind)
Denis Boyda
(ANL, MIT)
Gurtej Kanwar
(MIT)
Kyle Stuart Cranmer
(New York University (US))
Michael Albergo
(New York University)
Prof.
Phiala Shanahan
(Massachusetts Institute of Technology)
Sébastien Racanière
(DeepMind)
