Abstract
We develop a framework for generalized variational inference ininfinite-dimensional function spaces and use it to construct a method termedGaussian Wasserstein inference (GWI). GWI leverages the Wasserstein distancebetween Gaussian measures on the Hilbert space of square-integrable functionsin order to determine a variational posterior using a tractable optimisationcriterion and avoids pathologies arising in standard variational function spaceinference. An exciting application of GWI is the ability to use deep neuralnetworks in the variational parametrisation of GWI, combining their superiorpredictive performance with the principled uncertainty quantification analogousto that of Gaussian processes. The proposed method obtains state-of-the-artperformance on several benchmark datasets.