GAN-Based Priors for Quantifying Uncertainty in Supervised Learning

摘要

Machine learning and deep learning algorithms are frequently used in critical tasks where their output is used in high-stakes downstream applications. In these cases it is important to quantify the uncertainty in the predictions of these algorithms. Motivated by this, we present a novel learning-based Bayesian inference approach for quantifying uncertainty in a prediction. Our approach uses samples drawn from the joint distribution of measurements and predictions to train a generative adversarial network (GAN). Thereafter, given a noisy measurement, we use the generator component of the GAN as a prior in a Bayesian update. By reformulating the resulting high-dimensional posterior sampling problem to the low-dimensional latent space of the GAN, we are able to perform efficient Markov Chain Monte Carlo (MCMC) updates. We apply this approach to image classification and image inpainting problems in computer vision and to forward and inverse uncertainty quantification tasks arising in computational physics, and demonstrate how the ability to quantify uncertainty can be used to (a) detect samples that lie outside the distribution of the training samples, (b) quantify the confidence in the prediction, and (c) determine the subsequent measurement within an active learning strategy.