Core Concepts
Foundation models in physics data aim to improve performance and reduce training time.
Abstract
Foundation models are multi-dataset and multi-task machine learning methods.
They aim to improve physics performance and reduce training time.
OmniJet-α model demonstrates transfer learning between unsupervised and supervised tasks.
Tokenization strategies and transformer backbone are key components.
Generative model trained for jet physics shows good agreement with ground truth.
Transfer learning from generation to classification shows significant improvement in accuracy.
Stats
"Foundation models are multi-dataset and multi-task machine learning methods."
"The OmniJet-α model demonstrates transfer learning between an unsupervised problem (jet generation) and a classic supervised task (jet tagging)."
"The model was trained on three separate datasets: t → bqq′ only, q/g only, and q/g and t → bqq′ combined."
Quotes
"Foundation models for physics data would be significant: While machine learning models developed so far typically outperform classical approaches, available statistics for training these models is a constant issue."
"Foundation models will play an important role in reducing the computational burden in particle physics."