Understanding Event-Generation Networks via Uncertainties

Abstract

Following the growing success of generative neural networks in LHC simulations, the crucial question is how to control the networks and assign uncertainties to their event output. We show how Bayesian normalizing flow or invertible networks capture uncertainties from the training and turn them into an uncertainty on the event weight. Fundamentally, the interplay between density and uncertainty estimates indicates that these networks learn functions in analogy to parameter fits rather than binned event counts. Content 4 LHC events with uncertainties 18 5 Outlook 20 References 21

What's Anomalous in LHC Jets?

“…They allow access to the Jacobian and both directions of the mapping, linking density estimation in the physics and latent spaces in a completely controlled manner. We have used the flexible INN setup successfully for precision event generation [93,94], unfolding detector effects [95], and QCD or astro-particle inference [96,97].…”

Ephemeral Learning -- Augmenting Triggers with Online-Trained Normalizing Flows

“…[44], for a discussion of trainingrelated uncertainties using Bayesian normalizing flows see Refs. [45,46]. The properties of online training, specifically seeing every event independently and only once, are in tension with training generative models.…”

“…These weights are independent for each flow in the ensemble and are kept constant as the flow iterates over one buffer. We leave the further exploration of this ad hoc solution and alternative methods such as Bayesian flows [45,52,53] to the future.…”

Calomplification -- The Power of Generative Calorimeter Models

“…GANplification arises, intuitively, from the fact that neural networks work like classical parametric fits [4,5], and they are particularly effective when we want to interpolate in many dimensions. This feature is behind the success of the NNPDF parton densities [6] as the first mainstream ML-application in particle theory.…”