Fig. 6: Architecture and training curves of the generative adversarial network. | Communications Physics

Fig. 6: Architecture and training curves of the generative adversarial network.

From: Deep-learning-based decomposition of overlapping-sparse images: application at the vertex of simulated neutrino interactions

Fig. 6

a At training time, candidate kinematics are combined with random noise vectors and passed to the generator (highlighted in red), which transforms them into a synthetic image. The task of the discriminator (depicted in green) is to distinguish between true and synthetic images. The diagram illustrates the loss functions minimised during the training. b Training and validation curves display the performance of the different components within the GAN used for proton generation (similar curves are observed for other particles). The purple area denotes the average overlap between the GAN-generated and simulated distributions for five randomly selected parameter sets, each consisting of 10,000 real and synthetic events. On the right-hand side, we illustrate the true and GAN-generated distributions, highlighting their overlap for a randomly selected parameter set obtained from an arbitrary model checkpoint. It serves as an indicator of the overall performance of the GAN, with larger values indicating superior performance.

Back to article page