Unsupervised Training of Convex Regularizers Using Maximum Likelihood Estimation

Unsupervised learning is a training approach in the situation where groundtruth data is unavailable, such as inverse imaging problems. We present anunsupervised Bayesian training approach to learning convex neural networkregularizers using a fixed noisy dataset, based on a dual Markov chainestimation method. Compared to classical supervised adversarial regularizationmethods, where there is access to both clean images as well as unlimited tonoisy copies, we demonstrate close performance on natural image Gaussiandeconvolution and Poisson denoising tasks.