Abstract
Regression on function spaces is typically limited to models with Gaussianprocess priors. We introduce the notion of universal functional regression, inwhich we aim to learn a prior distribution over non-Gaussian function spacesthat remains mathematically tractable for functional regression. To do this, wedevelop Neural Operator Flows (OpFlow), an infinite-dimensional extension ofnormalizing flows. OpFlow is an invertible operator that maps the (potentiallyunknown) data function space into a Gaussian process, allowing for exactlikelihood estimation of functional point evaluations. OpFlow enables robustand accurate uncertainty quantification via drawing posterior samples of theGaussian process and subsequently mapping them into the data function space. Weempirically study the performance of OpFlow on regression and generation taskswith data generated from Gaussian processes with known posterior forms andnon-Gaussian processes, as well as real-world earthquake seismograms with anunknown closed-form distribution.