Inverse problems in image processing and computer vision are often solved using prior probability densities, such as spectral or sparsity models.  Machine learning now offers state-of-the-art solutions for most of these problems using artificial neural networks, which are typically optimized using nonlinear regression to provide direct solutions for each specific task.  As such, the prior probabilities are implicit, and intertwined with the tasks for which they are optimized.  I'll examine some properties of priors implicitly embedded in denoising networks, review methods for drawing samples from them. Extensions of these sampling methods enable the use of the implicit prior to solve any deterministic linear inverse problem, with no additional training, thus exploiting the power of supervised learning for denoising for use in a much broader set of problems. These methods rely on minimal assumptions, exhibits robust convergence over a wide range of parameter choices, and achieve state-of-the-art levels of unsupervised performance for deblurring, super-resolution, and compressive sensing.  They also offer opportunities for experimental probes of biological vision, and a potential basis for unsupervised learning in the brain.

Professor Simoncelli's talk will be Saturday, September 26 at 4:30pm.