Visual Image Priors in the Era of Machine Learning

Inverse problems in image processing and computer vision are often solved using prior probability densities, such as spectral or sparsity models. We also believe that biological visual systems, including those of humans, contain representations of such densities, allowing them to detect outliers and infer properties of the world from partial and noisy retinal signals. In recent years, machine learning has provided remarkably powerful models for these densities that can be learned, unsupervised, from data. These models offer dramatic improvements in solving traditional inverse problems, and in manipulating and generating visual stimuli for the study of biological visual systems. They also suggest potential means by which biological systems might learn and represent such densities.