Generative Adversarial Networks Acceleration for LHCb RICH Detector

Generative Adversarial Networks(GANs) are widely used today in various tasks of real data replication; however, strong quality performance of GANs is usually achieved through a significant increase of trainable parameters, which also increases computational complexity. This paper explores a way to accelerate a GAN used for particle generation for the LHCb RICH detector through compression of the model. The approach we use in this paper is called Knowledge Distillation; it allows us to train a shallow model with a smaller number of parameters using the initial heavyweight model as the teacher, without significant performance loss. The compressed models achieved in this project provide 37\% faster particle generation than the base model, with the number of trainable parameters reduced by 44\% while maintaining good score performance.