Statistics Seminar

Abstract: Data science tasks and statistical estimation often rely on optimization, yet existing optimization methods face two challenges for data science. First, optimization problems arising in data science have general geometric and noise properties which existing optimization methods are not designed to handle, which results in methods that are unreliable or impractical. Second, optimization methods are not able to readily exploit the emerging computing architectures driven by data science needs, which results in methods that are inefficient. We address these challenges by developing frameworks for optimization (and underlying linear algebra problems) that have rigorous convergence guarantees and that allow for methods that can be adapted to fully exploit specific computing environments.