Plenary talk solving linear inverse systems with graphcuts

Ill-conditioned linear inverse systems, which arise in many applications, are ambiguous problems that cannot be solved without some form of prior. Convex minimization methods can solve these problems efficiently, but only by making the unrealistic assumption that images are globally smooth. A more reasonable assumption, such as piecewise smoothness, results in an intractable optimization, that requires minimizing a highly non-convex function in a space with many thousands of dimensions. I will describe how graph cuts can be used to solve this problem, and demonstrate some applications to an important medical imaging problem. This is joint work with Ashish Raj (Cornell Radiology).