Topic：Optimization methods for solving structured high dimension problems
Speaker：Ph.D. Shaozhe Tao (University of Minnesota)
Time：2017-5-10 (Wednesday) 14:00
Place：G508
Abstarct: The talk will present my recent work in optimization methods for solving structured high dimension problems. In first part of the talk, we show all common first-order method, such as ISTA, FISTA, ADMM, coordinate descent, exhibit local linear convergence for LASSO problem. Using a spectral analysis, we show that, when close enough to the solution, FISTA slows down compared to ISTA, making it advantageous to switch to ISTA towards the end. In the second part, we propose a novel estimator for inverse covariance matrix with group structure. The problem can be efficiently solved via Frank-Wolfe method, leveraging chordal sparsity for scalability. Numerical results on synthetic and real datasets show significant improvement in sample complexity and performance.