News
CIDAS Colloquium
ML for efficient generation of particle scattering events
Dr. Timo Janßen (University of Göttingen, CIDAS)
In high energy physics, theory-based predictions come in the form of simulated scattering events. State of the art calculations require significant computation time of which a major part goes into the generation of valid configurations of particle momenta. In this talk I will show how generative models can reduce the computational load and in turn increase the physics reach. Based on an accessible introduction of the underlying physics, I will define the actual sampling problem as a statistical challenge. I will then show how flow-based models can contribute to solving this problem. Recently introduced simulation-free training methods make it possible to train these models at a reasonable cost and reveal interesting connections with diffusion / score-based generative models. Finally, I will show how optimal transport can increase the sampling speed by straightening the flow.
ML for efficient generation of particle scattering events
Dr. Timo Janßen (University of Göttingen, CIDAS)
In high energy physics, theory-based predictions come in the form of simulated scattering events. State of the art calculations require significant computation time of which a major part goes into the generation of valid configurations of particle momenta. In this talk I will show how generative models can reduce the computational load and in turn increase the physics reach. Based on an accessible introduction of the underlying physics, I will define the actual sampling problem as a statistical challenge. I will then show how flow-based models can contribute to solving this problem. Recently introduced simulation-free training methods make it possible to train these models at a reasonable cost and reveal interesting connections with diffusion / score-based generative models. Finally, I will show how optimal transport can increase the sampling speed by straightening the flow.