基于流形正则化框架和MMD的域自适应BLS模型

Domain Adaptive BLS Model Based on Manifold Regularization Framework and MMD

宽度学习系统(Broad learning system,BLS)作为一种基于随机向量函数型网络(Random vector functionallink network,RVFLN)的高效增量学习系统,具有快速自适应模型结构选择能力和高精度的特点.但针对目标分类任务中有标签数据匮乏问题,传统的BLS难以借助相关领域知识来提升目标域的分类效果,为此提出一种基于流形正则化框架和最大均值差异(Maximum mean discrepancy,MMD)的域适应BLS(Domain adaptive BLS,DABLS)模型,实现目标域无标签条件下的跨域图像分类.DABLS模型首先构造BLS的特征节点和增强节点,从源域和目标域数据中有效提取特征;再利用流形正则化框架构造拉普拉斯矩阵,以探索目标域数据中的流形特性,挖掘目标域数据的潜在信息.然后基于迁移学习方法构建源域数据与目标域数据之间的MMD惩罚项,以匹配源域和目标域之间的投影均值;将特征节点、增强节点、MMD惩罚项和拉普拉斯矩阵相结合,构造目标函数,并采用岭回归分析法对其求解,获得输出系数,从而提高模型的跨域分类性能.最后在不同图像数据集上进行大量的验证与对比实验,结果表明DABLS在不同图像数据集上均能获得较好的跨域分类性能,具有较强的泛化能力和较好的稳定性.

As an efficient incremental learning system based on random vector function-link network(RVFLN),broad learning system(BLS)has the characteristics of fast adaptive model structure selection and high precision.However,due to the lack of label data in target classification,the traditional BLS is difficult to improve the classification effect of target domain by using relevant domain knowledge.Therefore,a domain adaptive BLS(DABLS)model based on manifold regularization framework and maximum mean discrepancy(MMD)is developed to achieve cross-domain image classification of target domain under unlabeled condition.Firstly,the feature nodes and enhancement nodes of BLS are constructed to effectively extract features from the data of source domain and target domain.The manifold regularization framework is used to construct Laplacian matrix in order to explore the manifold characteristics of the target domain data and mine the potential information of the target domain data.Then the transfer learning method is used to construct the MMD penalty term between the source domain data and the target domain data to match the projection mean between the source domain and the target domain.The feature nodes,enhancement nodes,MMD penalty term and Laplacian matrix are combined to construct the objective function.Ridge regression analysis is used to solve the objective function to obtain the output coefficients,so as to improve the cross-domain classification performance.Finally,a large number of validation and comparative experiments are carried out on different image data sets,and the experiment results show that the DABLS can better achieve cross-domain classification on different image data sets,and has strong generalization ability and better stability.