Dr. Jinrae Kim visited CDSL and gave a seminar on “Universal convex approximation for learning objective function in amortized optimization”. He received his Ph.D. degree in aerospace engineering from Seoul National University, Seoul, Republic of Korea, and will begin his postdoctoral course in University of Illinois Urbana-Champaign.
One of Dr. Kim’s research interests is amortized optimization, which is a learning-based optimization framework to approximate the minimizer or the objective function of an optimization problem. In this seminar, Dr. Kim introduced his recent result that 1) proposes a specific neural network structure to obtain an objective function approximator for a non-convex optimization problem; 2) expresses the objective function approximator as a sum of a convex function and a residue in which the convex function shares the minimizer with the approximator. He emphasized that the proposed method converts the problem of finding the minimizer for the objective function approximator, which typically involves non-convex optimization, to an equivalent convex optimization problem. Additionally, the proposed neural network is proven to have high expressiveness in a sense that it is capable of approximating any continuous functions.
Since non-convex optimization problems frequently appear as a challenge in control theory, this seminar was very hopeful and informative for the researchers of CDSL. We wish success on Dr. Kim’s new journey as a postdoctoral candidate and look forward on collaborating with CDSL in the future.