Transfer learning for visual recognition and text categorization.
Date of Issue2012
School of Computer Engineering
Centre for Multimedia and Network Technology
In recent decades, transfer learning has attracted intensive attention from researchers and become a hot research direction in the field of machine learning. Different from traditional machine learning, transfer learning allows that the training and testing data can be from different domains (i.e., different data distributions and/or different feature spaces). This characteristic helps us to learn good classifiers for the domain of interest, where there have only a limited or even no labeled training data, by utilizing many existing data from other related sources. Because of this, transfer learning techniques have already been widely used in many areas such as machine learning, data mining, computer vision, etc. In this thesis, we propose several transfer learning frameworks, based on which a number of transfer learning are developed and applied for different real-world applications such as visual recognition and text categorization. Specifically, first we propose a domain transfer framework based on multiple kernel learning to minimize the data distribution mismatch between domains. Two methods are further developed under this framework to simultaneously learn a kernel function modeled by multiple kernel learning as well as a robust target classifier. We demonstrate the effectiveness of our proposed methods in the video concept detection and text classification tasks. Second, we present a visual event recognition framework for consumer videos by leveraging a large number of web videos, in which a pyramid matching method and a transfer learning method have been proposed to measure the distances between videos and cope with the data distribution mismatch between the consumer and web video domains, respectively. In the proposed transfer learning method, we define the target decision function as a linear combination of pre-learned classifiers and a perturbation function modeled by multiple kernel learning, such that we can better fuse the knowledge learned from multiple levels of a video and also different types of features, which helps learn a robust classifier. Third, we propose a domain-dependent regularization framework to handle the transfer learning problems where there exist multiple source domains. In this framework, a domain-dependent regularizer is defined based on a set of pre-learned classifiers by enforcing the smoothness that the target classifier shares similar decision values with the pre-learned classifiers on the target unlabeled samples. Furthermore, two methods are presented by incorporating least-squares SVM into the proposed framework. One of them employs a sparsity regularizer based on the $\epsilon$-insensitive loss. And the other one additionally makes use of the Universum regularizer defined on data from the source domains. Experimental results show the good performances of our proposed methods in the video concept detection and information retrieval tasks with multiple source domain settings.
DRNTU::Engineering::Computer science and engineering