Speaker
Description
The determination of parton distribution functions (PDFs) is core problem in hadron colliders physics, as PDFs model the initial state of the colliding partons. Traditionally, these distributions have been obtained by fitting a functional form to a subset of experimental data, making them susceptible to biases from these choices.
The NNPDF collaboration introduced an alternative approach in which the PDF is modeled using neural networks, removing the bias inherent to the choice of functional form and enabling a fully data-driven determination.
In this talk, I explore the power of ensemble regression to take this approach one step further by eliminating the choice of a specific machine learning architecture. By sampling from a space of possible models that can accommodate the collider data. This new approach, which incorporate the (hyper)parametric uncertainty directly into the determination has been made feasible by recent developments in hardware acceleration and opens the door to more robust extractions of PDFs.
| AI keywords | uncertainty estimation, ensemble regression, model sampling, data-driven models |
|---|
