| Machine learning has achieved great success in many fields with big data during the past decades,including computer vision,natural language processing,and smart medical and so on.The big data has brought opportunities as well as challenges for machine learning.The annotation of data is one of the challenges that widely exists.The data types are various,the environment of data generation is complex,and the speed of data updating is fast and so on,resulting in that only a few labeled data can be used in many real world applications and the generalization ability of learning model is poor.How to improve the classification accuracy with limited labeled samples,namely small size samples problem,becomes one of the hot topics in machine learning field.Active learning is an approach by the interaction between machine and human experts to address the small size samples problem.It selects the most informative samples for human labeling from large-scale data to build a high quality training data set.Then a strong generalization ability model can be trained with these high quality data.To improve the active learning performance,this thesis presents a systematic study on the models of samples selection with the combination of uncertain and representative information.The novel contributions are summarized as follows:1)For the classification task of single label data,this thesis proposes a heuristic active learning framework by combing the two-samples theory and uncertainty together.In this framework,it almost contains all the constraints that can be used to select the optimal samples.In this way,the poor adaptability of the related works is overcome,due to their data structure presentation with various assumptions.The active learning methods that are designed based on this framework can select the samples with uncertainty and representativeness simultaneously.2)For the classification of multi-label data,this thesis proposes a robust multi-label active learning model based on maximum correntropy criterion to address the problem that the abnormally low relevant label impacts on the accurate measure of uncertainty and representativeness.The proposed model mainly depends on the bounded property of the maximum correntropy criterion,which can restrain the influence of low relevant labels to measure the two criteria and make the strong relevant labels dominate the measures.3)To address the shortage of semi-supervised information in the combination of semi-supervised learning and active learning,this thesis proposes a semi-supervised active learning method by integrating the discriminative and representative information together.It mainly uses the labeled data that is before and after updated in the active learning loop to build several classification models and clustering models.Then the discriminative samples can be obtained by assigning pseudo labels with high confidence based on the classification models,and the representative samples can be obtained by assigning pseudo labels based on clustering models.While the samples that can not be assigned with pseudo under the mechanism are regarded as candidates for active learning.By this way,the discriminative and representative information in labeled data are greatly enhanced,and the active learning performance is efficiently improved.4)For the problem that the uncertainty and representativeness criteria are usually not unified in the combination of domain adaptation and active learning,this thesis proposes multiple domains active learning with domain adaptation.It embeds the source data and target data in an active learning framework which combines the uncertainty and representativenss together to select the most informative samples from target domain.Meanwhile,the distribution of source domain can be constantly adjusted in active learning to make the source data be suitable for the target classification task,and the accuracy is improved for the cross-domain classification. |
