| 英文摘要 | In recent years, with the increasing network bandwidth and the development of sensing and storage devices, data collection is getting cheaper, while data processing and analysis are facing unprecedented challenges. How to utilize the ever-increasing data to adapt recognition models has becoming an important problem, which needs to be solved urgently. Research on online adaptive learning has attracted much attention from both the academia and industry in the last decades. When human beings deal with massive amounts of data, we continue to acquire new knowledge to adapt to environmental changes with little supervision. With attempt to simulate such human cognition mechanism, we study several online paradigms, i.e., online semi-supervised learning,
active learning and continuous adaptation. A number of effective learning methods are proposed and have been successfully applied to classification tasks. The main contributions of the thesis are summarized as follows:
• To reduce the labeling burden in data stream, we propose a novel online semisupervised algorithm based on prototype learning, called OSS-LVQ. Here we assume that data points clustered together are likely to have the same class label (i.e., each class is divided into several clusters). Supervised prototype-based classifier model and unsupervised prototype-based cluster model are learned in a unified framework, where the prototypes are shared by supervised and unsupervised learning. The method can utilize both labeled samples and unlabeled samples in data stream to improve the classification performance. We conduct experiments on handwriting images, natural scene images and text datasets, and verify the effectiveness of the method in respect of classification accuracy and time efficiency.
• To explore the continuous learning problem in an open environment, we propose a semi-supervised class-incremental learning method based on learning vector quantization, called EvolvingProto. This method firstly combines probabilistic prototype model with a from-coarse-to-fine learning stategy to detect novel classes from unlabeled samples in data stream, and then a small amount of samples from novel classes are chosen to be labeled based on prototype clustering, finally the capability to classify novel classes are integrated into the model by memorizing critical samples. The experimental results show that the proposed method has capability to discover novel classes from unlabeled data stream, and realize class-incremental learning with limited labeled samples.
• To tackle the continuous change of sample style in the data stream, we propose an incremental adaptive learning vector quantization, called CILVQ. In this method, unified transfer matrix is used to model style for all samples, and then adjusted according to style changes. Finally, test samples can be classified in the style-free space. To reduce the cost of the labeling process, we exploit the active learning to update classifier model. The experimental results on handwritten character dataset show that the
proposed method can improve the classification performance with the local style consistency. Meanwhile, classifier model automatically adapt to writing style when the writers change. |
修改评论