AI and the World's Hard Problems · Physics · arXiv · 2026-10-23
Faithful Uncertainties in ML-Based Inverse Problems: GPs, NNs and NTKs for the Triplet PDF
Machine-learning solutions of physics inverse problems need faithful uncertainties, not just good central values. Using a closure-test setup for reconstructing the non-singlet triplet distribution T3 from deep-inelastic scattering data, the authors compared Gaussian processes, neural-network ensembles and empirical neural tangent kernels. Methods with similar central values gave very different uncertainties, controlled by the likelihood's covariance treatment, the functional prior and the training stage, and calibration depended on predictive correlations rather than pointwise width.