Abstract
Artificial intelligent and machine learning technologies have already achieved significant success in classification, regression and clustering. However, many machine learning methods work well only under a common assumption that training and test data are drawn from the same feature space and the same distribution. A real world applications is in sports footage, where an intelligent system has been designed and trained to detect score-changing events in a Tennis single match and we are interested to transfer this learning to either Tennis doubles game or even a more challenging domain such as Badminton. In such distribution changes, most statistical models need to be rebuilt, using newly collected training data. In many real world applications, it is expensive or even impossible to collect the required training data and rebuild the models. One of the ultimate goals of the open ended learning systems is to take advantage of previous experience/ knowledge in dealing with similar future problems. Two levels of learning can be identified in such scenarios. One draws on the data by capturing the pattern and regularities which enables reliable predictions on new samples. The other starts from an acquired source of knowledge and focuses on how to generalise it to a new target concept; this is also known as transfer learning which is going to be the main focus of this thesis. This work is devoted to a second level of learning by focusing on how to transfer information from previous learnings, exploiting it on a new learning problem with not supervisory information available for new target data. We propose several solutions to such tasks by leveraging over prior models or features. In the first part of the thesis we show how to estimate reliable transformations from the source domain to the target domain with the aim of reducing the dissimilarities between the source class-conditional distribution and a new unlabelled target distribution. We then later present a fully automated transfer learning framework which approaches the problem by combining four types of adaptation: a projection to lower dimensional space that is shared between the two domains, a set of local transformations to further increase the domain similarity, a classifier parameter adaptation method which modifies the learner for the new domain and a set of class-conditional transformations aiming to increase the similarity between the posterior probability of samples in the source and target sets. We conduct experiments on a wide range of image and video classification tasks. We test our proposed methods and show that, in all cases, leveraging knowledge from a related domain can improve performance when there are no labels available for direct training on the new target data.