Evaluating Generative Adversarial Networks for particle hit generation in a cylindrical drift chamber using Fréchet Inception Distance

Abstract

Abstract We use Fréchet Inception Distance (FID) measured in the latent spaces of pre-trained, fine-tuned and custom-made inception networks to evaluate Generative Adversarial Networks (GANs) developed by the COherent Muon to Electron Transition (COMET) collaboration to generate sequences of background hits in a Cylindrical Drift Chamber (CDC). We validate the convergence of the GANs' training and show that the use of self-attention layers reduces FID. Our method enables the use of FID as an evaluation metric even when an application-specific inception network is not readily available, making it transferable to other GAN applications in High Energy Physics.