Speaker
Description
The calorimeter in Large Hadron Collider (LHC) experiments measures particle energy by tracking showers from collisions. Describing these processes requires precise simulation methods, such as the Geant4 toolkit. Recently, generative models have emerged as a faster alternative based on different Machine Learning (ML) architectures, such as Diffusion and Variational Autoencoders.
The training of ML models is predominantly carried out using Python frameworks, primarily PyTorch and TensorFlow. In order to determine how mature ML development is using Julia, a denoising diffusion model, CaloDiffusion, was chosen to be implemented and trained with Flux.jl. On top of technical details, this talk also covers benchmarks of both implementations and analysis of performance using GPU profiling.
