Understanding Event-Generation Networks via Uncertainties

TL;DR

This paper explores how Bayesian normalizing flow networks can quantify uncertainties in generated events, revealing their function-learning nature akin to parameter fitting rather than traditional histogram-based methods.

Contribution

It demonstrates that invertible neural networks can effectively capture and interpret uncertainties in event generation, advancing control and reliability in LHC simulations.

Findings

Uncertainty estimates correlate with training data variability

Networks learn functions similar to parameter fits

Uncertainty propagation to event weights

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.