LHC analysis-specific datasets with Generative Adversarial Networks
- Hashemi, Bobak et al
- arXiv:1901.05282
 | Distribution of six quantities computed from the 18 considered features of the $Z\to \mu\mu$ dataset. |
 | Pairwise correlation matrices of features for target (left) and generated (right) samples, the diagonal terms have been set to 0 for aesthetics. The pdf for this model is not expected to be a high dimensional Gaussian, so the correlation matrix must be taken with a grain of salt. |
 | Training history for a typical network training on the reduced (top) and full (bottom) dataset. Left: Discriminator and generator loss as a function of the training epoch. Center: Mean and standard deviation of invariant masses calculated from generated samples. Right: Sum of Kolmogorov-Smirnov test statistics across all marginal distributions. |
 | Network architectures for the discriminator (top) and generator (bottom). |
 | Distribution of the selected features in the target dataset, Drell-Yan to dimuon events generated with {\tt PYTHIA8} and reconstructed with the CMS detector simulation in {\tt DELPHES} |
 | Left: Z-boson transverse momentum in the target and generated datasets. Right: Comparison of the $E_T^{\mathrm{miss}}$ distributions in the low-pileup and high-pileup regime, for the true and generated events. |
 | Comparison between the distribution of input and GAN-generated quantities, and for the $\mll$ distribution derived from them, obtained removing the two $\mll$-related terms from the loss function in Eq.~(\ref{eq:gen_loss}). |
 | Comparison between the distribution of input and GAN-generated quantities, and for the $\mll$ distribution derived from them and entering the minimized loss function defined in Eq.~(\ref{eq:gen_loss}). |
 | Left: Z-boson transverse momentum in the target and generated datasets. Right: Comparison of the $E_T^{\mathrm{miss}}$ distributions in the low-pileup and high-pileup regime, for the true and generated events. |
 | Comparison of Z-boson transverse momentum plotted against the leading-jet $\pt$, for target (left) and generated (right) samples. |
 | Training history for a typical network training on the reduced (top) and full (bottom) dataset. Left: Discriminator and generator loss as a function of the training epoch. Center: Mean and standard deviation of invariant masses calculated from generated samples. Right: Sum of Kolmogorov-Smirnov test statistics across all marginal distributions. |
 | Network architectures for the discriminator (top) and generator (bottom). |
 | Comparison target and input distributions for a set of quantities derived from the 17 quantities returned by the GAN. |
 | Comparison of the target distributions to those generated by the generator model for the full dataset. |