Speaker
Description
With the vast amount of data already produced by the LHC and the upcoming HL-LHC runs, reliable and precise uncertainty quantification for numerical tools, in particular neural networks, is essential. We present a systematic study of uncertainty estimation using various machine learning approaches, including neural networks, Gaussian processes, and neural tangent kernels.
As an underlying dataset, we use the T3 contribution to parton distribution functions (PDFs), based on datasets already employed in NNPDF, to provide a controllable yet well-understood framework for our analysis. We investigate how uncertainty estimates depend on the modelling strategy and its underlying assumptions, including the incorporation of physics knowledge. In addition, we test all methods in extrapolation regions and assess the trustworthiness of their uncertainty estimates.
For practical validation, we compare the T3 model predictions obtained with different numerical setups to deep-inelastic scattering data from HERA, providing a phenomenological benchmark for the predictions and their uncertainties. Our results highlight the strong method dependence of ML-based uncertainty estimates and underscore the need for careful validation when applying these techniques in collider phenomenology.