< Back to previous page

Project

Unsupervised and semi-supervised learning using kernel methods.

This research proposal is aimed at addressing key challenges and open problems in unsupervised and semi-supervised learning using kernel-based methods. The least squares support vector machine (LS-SVM)framework is considered as a core methodology. This family of kernel methods is cast in a constrained optimization setting. The primal problem is expressed in terms of high dimensional feature maps while the dual problem is expressed in terms of positive definite kernels. The main advantages of using this methodology are the possibility to extend the core model by adding additional constraints into the primal problem and the out-of-sample extension, which allows the trained model to be applied to new data. The research objectives that will be investigated are summarized in the following aspects: implementing new model selection and kernel design criteria, incorporating prior information into the analysis, developing different approaches for sparseness to handle large-scale data, exploring new clustering algorithms and extending the current models to work with time-dependent data. The methodologies to tackle these problems include using the out-of-sample extension and PageRank algorithms to achieve sparseness, adding constraints to incorporate prior information and structure, together with using special properties of the solutions to perform model selction and kernel design.
Date:1 Oct 2009 →  30 Sep 2010
Keywords:Optimization, Model selection, Support vector machines, Kernel methods, Semi-supervised learning, Unsupervised learning
Disciplines:Modelling, Multimedia processing