Conformal Uncertainty Quantification Guarantees for Neural Operators
The paper gives neural-operator predictions conformal bands with explicit domain-coverage guarantees.
Stent and Boullé build a split conformal method that wraps a calibrated pointwise band around a neural operator’s output. The guarantee is that the true solution is covered on at least a chosen fraction of the evaluation domain, with a stated probability over calibration and test inputs. Their proof applies to measurable residual fields on arbitrary probability spaces, including both continuum domains and fixed grids. In Darcy flow and Navier-Stokes experiments, the calibrated bands keep target coverage while coming out tighter than existing corrections. ArXiv · AI/CL/LG's note
Stent and Boullé build a split conformal method that wraps a calibrated pointwise band around a neural operator’s output. The guarantee is that the true solution is covered on at least a chosen fraction of the evaluation domain, with a stated probability over calibration and test inputs. Their proof applies to measurable residual fields on arbitrary probability spaces, including both continuum domains and fixed grids. In Darcy flow and Navier-Stokes experiments, the calibrated bands keep target coverage while coming out tighter than existing corrections. ArXiv · AI/CL/LG's note
score 4