Research On Augmented Class Learning

In recent years,the learning paradigm oriented to open and dynamic environment has gradually attracted attention from various fields,where class-incremental learning(C-IL)is one of the hottest research topics.In C-IL,there are unseen classes in test set during the training phase,therefore,the model needs not only to correctly classify the instances of seen classes,but also to be able to automatically detect unseen classes.This problem is formally defined as learning with augmented class(LAC).In recent years,LAC problem has attracted an increasing number of attention and some effective algorithms have been proposed and achieved satisfactory performance.However,the research on LAC problem is still in its preliminary stage,and there are still plenty of problems in LAC problem that have not been well studied,including:(1)lacking of supervision information of unseen classes;(2)imbalanced class distributions among seen classes are widely-existing in the real-world applications of augmented class learning;(3)the scattered distribution of unseen classes will further increase learning difficulties.To tackle these three challenges,two approaches on the basis of the LACU framework are proposed in this thesis:Lacking of supervision information of unseen classes is an essential difficulty of LAC problem,and the imbalanced class distributions among seen classes will further increase learning difficulties.Hence,a label confidence propagation based novel approach LCP is proposed to solve the imbalanced augmented class learning problem.LCP enlarges the labeled training data set by estimating class labels of unlabeled data.Thus on one hand,the supervision information of unseen classes can be estimated,which will greatly reduce the difficulty of learning the concepts of unseen classes.And on the other hand,more training data is available to represent class concepts more sufficiently,which will alleviate the damage of class-imbalance.Results on abundant experiments clearly validate the effectiveness of the proposed LCP and its robustness to high imbalance ratio and high level of open environment.And the more unlabeled data available,the better the advantages of LCP.When multiple unseen classes are treated as one novel class,the class actually contains multiple sub-class concepts,and its data distribution in feature space is usually very scattered,which will increase the learning difficulties.Accordingly,a feature selection based novel approach LACU-FS is proposed.LACU-FS aims to construct a feature subspace that enables all unseen classes to assemble as much as possible without destroying the separability of each seen class,thus abating the difficulty of model identification for unseen classes.LACU-FS obtains the estimated samples of unseen classes by estimating class labels of unlabeled data.Then,for any combination of two classes in the label space,Random Forest algorithm is used to select the most discriminative feature set.The expected feature subspace can be obtained by merging the above-described feature sets.Results on abundant experiments clearly validate the effectiveness of the proposed LACU-FS,and the more complex the data,the better LACU-FS can demonstrate its advantages.The research background,difficulties of augmented class learning and the contribution of this thesis are introduced in chapter 1.Chapter 2 and chapter 3 introduces the proposed augmented class learning method LCP and LACU-FS,respectively.Finally,chapter 4 concludes the thesis and looks forward to the future research directions.