Inside Our ProgramProgram Events
Events
Upcoming Events
Events filtered by
-
Apr19
EVENT DETAILS
more lessVivak Patel, Ph.D.
University of Wisconsin -- Madison
Title: Global Convergence and Stability of Stochastic Gradient Descent
Abstract: Stochastic Gradient Descent (SGD) is one of the most important optimization algorithms for data science, and it is widely applied to optimize highly non-convex functions with equally complicated noise models. As a result, understanding whether SGD will actually find a solution to such problems is important to practitioners, and a first step towards this understanding is to study whether SGD is globally convergent. While previous work has studied the global convergence of SGD, such works have made assumptions that are often restrictive in practical settings. Here, we provide examples that demonstrate this restrictiveness, and provide assumptions that are more realistic. Under these more realistic assumptions, we prove that either SGD diverges or converges to a stationary point with probability one. We then move on to study what happens when SGD diverges, and we show that, under an additional assumption, SGD remains stable---that is, the objective function cannot diverge.
Bio: Vivak Patel is an assistant professor of statistics at the University of Wisconsin -- Madison. His research focuses on problems at the intersection of statistics and computing, which includes developing novel computational methods for statistical problems and developing novel statistical methods for computational methods.
TIME Tuesday, April 19, 2022 at 11:00 AM - 12:00 PM
LOCATION ITW 1.350, Ford Motor Company Engineering Design Center map it
CONTACT Agnes Kaminski a-kaminski@northwestern.edu EMAIL
CALENDAR Department of Industrial Engineering and Management Sciences
