Uncertainties associated with GAN-generated datasets in high energy physics

Matchev, Konstantin T.; Roman, Alexander; Shyamsundar, Prasanth

doi:10.48550/arxiv.2002.06307

Cited by 5 publications

(6 citation statements)

References 43 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2 A related question is the statistical power of the examples generated from G. See Ref. [70] and [71] for discussions, and the latter paper for an empirical demonstration of this topic. 3 We are neglecting the contribution to the variance from f which should also be estimated from the sample.…”

Section: B Statistical Properties Of Weighted Examplesmentioning

confidence: 99%

DCTRGAN: improving the precision of generative models with reweighting

Diefenbacher¹,

Eren²,

Kasieczka³

et al. 2020

J. Inst.

View full text Add to dashboard Cite

Significant advances in deep learning have led to more widely used and precise neural networkbased generative models such as Generative Adversarial Networks (Gans). We introduce a post-hoc correction to deep generative models to further improve their fidelity, based on the Deep neural networks using the Classification for Tuning and Reweighting (Dctr) protocol. The correction takes the form of a reweighting function that can be applied to generated examples when making predictions from the simulation. We illustrate this approach using Gans trained on standard multimodal probability densities as well as calorimeter simulations from high energy physics. We show that the weighted Gan examples significantly improve the accuracy of the generated samples without a large loss in statistical power. This approach could be applied to any generative model and is a promising refinement method for high energy physics applications and beyond.

show abstract

Section: B Statistical Properties Of Weighted Examplesmentioning

confidence: 99%

DCTRGAN: improving the precision of generative models with reweighting

Diefenbacher¹,

Eren²,

Kasieczka³

et al. 2020

J. Inst.

View full text Add to dashboard Cite

show abstract

“…Note that although the estimates of F, U, and σ syst /σ stat from a preliminary dataset will not be accurate up to the sensitivity offered by the full MC dataset [50], they will be sufficient for projecting, with sufficient accuracy, the sensitivity of the experiment under different sampling distributions.…”

Section: Oasis For Analysis Variables 31 Groundworkmentioning

confidence: 99%

“…This condition needs to only be satisfied almost everywhere for optimal estimation of F 4. An alternative approach, motivated by recent advances in Machine Learning (ML), replaces the exact distribution f with an (approximate) ML regressor trained on a small sample of events produced from f[47][48][49], leading to an increase in the rate but not necessarily the precision of the simulation[50]. These methods use points in the phase space for which f is computed, and not events sampled from f .…”

mentioning

confidence: 99%

OASIS: Optimal Analysis-Specific Importance Sampling for event generation

Matchev,

Shyamsundar

2020

Preprint

Self Cite

View full text Add to dashboard Cite

We propose a technique called Optimal Analysis-Specific Importance Sampling (OASIS) to reduce the number of simulated events required for a high-energy experimental analysis to reach a target sensitivity. We provide recipes to obtain the optimal sampling distributions which preferentially focus the event generation on the regions of phase space with high utility to the experimental analyses. OASIS leads to a conservation of resources at all stages of the Monte Carlo pipeline, including full-detector simulation, and is complementary to approaches which seek to speed-up the simulation pipeline. Contents

show abstract

“…A seemingly straightforward and intuitive answer to this question is as many examples as were used for training, because the network does not add any physics knowledge [39]. However, there are reasons to think that a generative model actually contains more statistical power than the original data set.…”

Section: Introductionmentioning

confidence: 99%

GANplifying Event Samples

Butter,

Diefenbacher,

Kasieczka

et al. 2020

Preprint

View full text Add to dashboard Cite

A critical question concerning generative networks applied to event generation in particle physics is if the generated events add statistical precision beyond the training sample. We show for a simple example with increasing dimensionality how generative networks indeed amplify the training statistics. We quantify their impact through an amplification factor or equivalent numbers of sampled events.

show abstract

Uncertainties associated with GAN-generated datasets in high energy physics

Cited by 5 publications

References 43 publications

DCTRGAN: improving the precision of generative models with reweighting

DCTRGAN: improving the precision of generative models with reweighting

OASIS: Optimal Analysis-Specific Importance Sampling for event generation

GANplifying Event Samples

Contact Info

Product

Resources

About