Soutenance de thèse Huyen Do

Mme Huyen Do soutiendra, en vue de l'obtention du grade de docteur ès sciences, mention informatique, sa thèse intitulée

A unified framework for SVM, Multiple Kernel Learning and Metric Learning

Abstract :

Kernel methods have become a popular learning paradigm in machine learning in the last decades. However, their performance strongly depends on the choice of kernel. Therefore, kernel learning becomes a crucial problem for all kernel-based methods, in which the central question is how to choose an appropriate kernel. The Multiple Kernel Learning (MKL) framework has been developed recently to tackle efficiently the kernel learning problem. In this framework, the learned kernel is a combination of some given basic kernels. Becoming one of the main streams in Machine Learning, MKL has attracted a considerable attention in the last ten years. It is not only a powerful framework to learn kernels, but also an efficient tool for learning with multiple sources and for feature selection. By exploiting the sparse nature of l1 norm, MKL can automatically select the appropriate data sources or groups of features or at the limit, do feature selection. In this thesis, we will focus on two applications of the MKL framework: kernel learning and feature selection. We will develop novel algorithms to tackle the problems of kernel learning and feature selection.

Kernel methods are one of the two main branches of similarity-based learning, a longstanding research topic in Machine Learning. The second main branch of similarity-based learning is distance-based methods, which have even a longer history than kernel methods. Learning a good similarity function, i.e a good kernel or a good distance metric plays a crucial role in the success of similarity based methods. In the last decade, together with kernel learning, distance metric learning have also grown with a fast pace in machine learning. However, metric learning is mostly limited to linear transformations. In this thesis, we will develop the MKL framework in the context of metric learning, combining linear and nonlinear transformations in a single learning problem, thus, extending the representation power of metric learning.

Both kernel-based and distance-based methods are based on the similarity of pairwise instances. Thus, kernel learning and metric learning are two sides of the same coin. While kernel learning learns an optimal kernel for kernel-based methods, metric learning learns an optimal metric for distance-based methods. However, the link between two learning paradigms is still not clear and not yet carefully studied. Kernel learning and metric learning algorithms have been developed separately as two distinguish lines of research, with almost no connection between them. In this thesis, we will establish the relation between kernel-based and metric learning methods, building a unified framework for SVM, MKL and metric learning. Unifying learning algorithms is an important task in Machine Learning, which gives more insights to the existing algorithms, deepening the understanding of them and which opens room to improve them, as well as to develop new algorithms. Inspired by this unified framework and by the relationship among the learning methods, we develop several novel learning algorithms which have proved to be efficient in several benchmark datasets.

Date: Lundi 22 octobre 2012 à 14h00

Lieu: Battelle bât. D - Auditoire D185

5 octobre 2012

À la Une