Speaker
Description
The complexity of jet tagging and regression tasks has driven much of the ML4HEP developments in the past decade. From this wave of research, transformers have emerged as a general-purpose architecture for particle physics, which additionally benefit from the scale of our massive datasets. This talk reviews recent developments in jet tagging and scaling studies that allow the prediction of optimal hyperparameters given a dataset size and/or compute budget. Transformers are naturally flexible for multimodal data, and we also review transformer variants used in earlier stages of the reconstruction pipeline, such as tracking and particle flow. Any ML algorithm is only as good as the data we provide. We summarise by highlighting how generative ML models can also be used to calibrate taggers or provide more ML training data, either through background estimation methods for hadronic final states or fast simulation. Across the board, these developments amplify the discovery potential of the LHC in our quest to understand the Standard Model and what might lie beyond.