The Landscape of Several Non-convex Statistical Learning Problems
Time: 2017-09-11
Published By: Xiaoni Tan
Speaker(s): Song Mei (Stanford University)
Time: 10:00-11:00 September 27, 2017
Venue: Room 9, Quan Zhai, BICMR
For general non-convex optimization problems, local search algorithms such as gradient descent and trust region method could stuck at bad local minima and fail to find the global minimum. However in practice, local search algorithms work very well on a handful of non-convex statistical learning problems. To fill this gap between theory and practice, we propose to study the landscape (global geometry) of these statistical models. In this talk, I will go over several non-convex statistical models, as well as several powerful tools such as uniform convergence and Kac-Rice formula. I will show how these tools characterize the nice global geometry of these statistical models, and therefore local search algorithms work well provably.
