Department of Mathematics
Geometric Landscape Analysis of Some Non-Convex Optimizations
Non-convex optimization is one of the most powerful tools for solving scientific and engineering problems. Many important problems, from traditional low-rank approximation in numerical linear algebra to modern deep neural network training in machine learning, are formulated as non-convex optimization problems. For many of such problems, simple algorithms, such as alternating minimization and gradient descent, often work remarkably well, despite possible local minima caused by non-convexity. The success of such simple algorithms remains a mystery for many problems with practical applications.
One way to examine why these simple algorithms perform successfully for non-convex optimization is to study the landscape of the objective functions by characterizing and locating their critical points. Recently, there has been a stream of research dedicated to the landscape analysis of non-convex functions arising in high-dimensional data analysis with low-dimensional models, including blind deconvolution, dictionary learning, tensor completion, phase synchronization, deep neural networks, low-rank matrix recovery, and phase retrieval. Those results revealed that there are no poor local minima of those non-convex functions. Consequently, any algorithm finding a local minimum will give a good solution. Therefore, the interested problems can be efficiently and effectively solved with a theoretical guarantee by a large class of algorithms.
In our research project, we investigated landscape of non-convex optimization arising from phase retrieval, which is a fundamental problem in many imaging techniques such as X-ray crystallography, transmission electron microscopy, and coherent diffractive imaging. We want to obtain a high-resolution image by using a limited number of samples. We have shown that, with a nearly optimal sample size, a family of non-convex optimizations for phase retrieval can have a benign landscape --- all local minima are global and all other critical points have negative directional curvatures. Combined with algorithms that are guaranteed to find a local minimum, phase retrieval problems can be efficiently solved with a theoretical guarantee using many optimization algorithms. Our results not only provide a comprehensive understanding on why simple non-convex algorithms usually find a global minimizer for phase retrieval, but also will be useful for developing new efficient algorithms with a theoretical guarantee by applying algorithms that are guaranteed to find a local minimum.
(852) 3469 2248
- Professor, Department of Mathematics
- Director of Image Sensing and Intelligent Robot System Lab