Prof. Xiaolin Huang
Project Description and Objectives
Kernel methods, which implicitly maps data into feature spaces with high, even indefinite, dimensions, are very important in machine learning and have been widely applied in many fields. In the recent years, the success of deep learning implies that enhancing the flexibility with the support of big data is promising to improve machine learning performance. The route is also applicable to advancing kernel methods, which are traditionally restricted to shallow structures.
In this project, we will investigate serval key issues for advanced kernel methods. First, it is necessary to design deeper structure, e.g., with several nonlinear layers, and develop the corresponding training methods. Second, making kernels flexible usually violate positive definiteness condition, that is usually required by classical kernels, and investigation on indefinite kernel methods is desirable. Third, flexible kernels need to admit value-defined matrices, for which out-of-sample extension technique is necessary.
The objectives of this project consist of:
1) Novel kernel methods in one of the three topics: deep kernel/indefinite kernel/out-of-sample extension;
2) Toolbox for the developed techniques.
Basic knowledge on machine learning.
Programming skills on Matlab, Python, C.
Develop novel machine learning methods based on flexible kernels.
Establish and release toolbox for the developed methods.