Open-Ended Lifelong Learning Based On Self-organizing Graph

The learning process of human is“lifelong”and“open-ended”.On the one hand,human can keep acquiring new knowledge from the novel instances without forgetting the knowledge that he or she has already learnt.On the other hand,human can also make full use of the previous knowledge to address new questions or learn new things.In this paper,a novel machine learning paradigm,namely the Open-Ended Lifelong Learning(OELL),is proposed to simulate the manner of how human learns.Different from other traditional machine learning paradigms,in OELL labeled and unlabeled instances could be learnt incrementally at the same time,novel categories could be learnt from new samples,and prediction could be made at any time while learning.In the era of big data when new data samples are created unceasingly,this learning paradigm will hopefully have some application potential in the fields of data mining,objects recognition and intelligent robots,etc.In this paper we will address the proposed OELL problem in both unsupervised and semi-supervised learning environments respectively.In the unsupervised learning environment,we propose learning method based on Self-organizing incremental neural network and Density Peaks Analysis(SDPA).The algorithm can cluster the streaming data in the online manner,and generate new clusters incrementally to represent the novel distribution of new data when the data distribution changes.To handle the more complex learning environment,namely the semi-supervised learning,we propose the Self-organizing Incremental Graph(SOIG)algorithm.The algorithm is capable of learning both labeled and unlabeled instances incrementally,building the graph model to represent the knowledge,and acquiring the knowledge of novel categories from new samples.We use the experiments on artificial and real data sets,and the application systems to examine the lifelong learning ability and the open-ended property of the proposed algorithms.Analysis about the effectiveness of the algorithms is presented in the experimental section.The main contribution of this paper includes:1.A novel machine learning paradigm,the Open-Ended Lifelong Learning,is pro-posed.2.The SDPA algorithm is designed to address the OELL problem in the unsuper-vised learning environment.Meanwhile,a image segmentation method based on the clustering of pixels is achieved to examine the effectiveness of SDPA.3.The SOIG algorithm is proposed to solve the OELL learning problem in the semi-supervised learning environment.4.Based on SOIG,an open-ended lifelong learning system for the moving instruc-tions of the robot is designed.