Coordinate Descent Methods Beyond Separability and Smoothness
Explore coordinate descent methods for optimization problems that go beyond traditional separability and smoothness constraints in this 30-minute talk by Ion Necoara at the Erwin Schrödinger International Institute for Mathematics and Physics. Delve into techniques for handling nonseparable and nonsmooth objective functions, including random coordinate proximal gradient methods and smooth approximation frameworks. Learn about the scalability of these algorithms through local approximation models along random subspaces. Examine the worst-case complexity analysis for both convex and nonconvex settings. Discover the practical applications of these methods in areas such as smallest eigenvalue problems, matrix factorization, and support vector machine classification.