Abstract
Deep Learning (DL) methods have been transforming computer vision withinnovative adaptations to other domains including climate change. For DL topervade Science and Engineering (S&E) applications where risk management is acore component, well-characterized uncertainty estimates must accompanypredictions. However, S&E observations and model-simulations often followheavily skewed distributions and are not well modeled with DL approaches, sincethey usually optimize a Gaussian, or Euclidean, likelihood loss. Recentdevelopments in Bayesian Deep Learning (BDL), which attempts to captureuncertainties from noisy observations, aleatoric, and from unknown modelparameters, epistemic, provide us a foundation. Here we present adiscrete-continuous BDL model with Gaussian and lognormal likelihoods foruncertainty quantification (UQ). We demonstrate the approach by developing UQestimates on `DeepSD', a super-resolution based DL model for StatisticalDownscaling (SD) in climate applied to precipitation, which follows anextremely skewed distribution. We find that the discrete-continuous modelsoutperform a basic Gaussian distribution in terms of predictive accuracy anduncertainty calibration. Furthermore, we find that the lognormal distribution,which can handle skewed distributions, produces quality uncertainty estimatesat the extremes. Such results may be important across S&E, as well as otherdomains such as finance and economics, where extremes are often of significantinterest. Furthermore, to our knowledge, this is the first UQ model in SD whereboth aleatoric and epistemic uncertainties are characterized.