Research On Domain Adaptation Based On Deep Learning

In visual fields,it is expensive to collect sufficient labeled data.The generalization ability of standard supervised learning methods can not meet the actual demand when labeled data is scarce.As an emerging machine learning method,domain adaptation learning aims at solving the problem of target domain classification with little or no labels by making use of labeled source domain to train a classifier.At present,there exist three critical issues and challenges: negative-transfer,under-fitting and under-adaptation,which bad transfer results stem from.A more challenging task is that the source and target domain data is in heterogeneous feature spaces,which makes transfer more difficult.Based on above problems,the main contents of this paper are as follows:Firstly,for homogeneous domain adaptation,a domain adaptation network based on autoencoder(DANA)is proposed in order to learn effective features and minimize the distribution discrepancy between domains to improve the under-adaptation problem.First,the source and target samples are subjected to two-layer coding and decoding operations respectively to learn more efficient feature expression by minimizing the reconstruction error;Then,the maximum mean discrepancy criterion is used to match the marginal and conditional distribution in the feature extraction layer and classification layer respectively to minimize the distribution discrepancy,and the label information of source domain samples is encoded by a softmax classifier to improve the classification performance;Finally,the gradient descent method is used to learn the network parameters,and the prediction of the unlabeled samples in the target domain is completed according to classifier output.Secondly,for homogeneous domain adaptation,it will lead to under-fitting and under-adaptation simultaneously if the learned model can't adequately describe probability distribution of the predictive data.Besides,the simple graph regularization can't avoid the problem of negative-transfer completely.Aiming at this problem,a domain adaptation network based on hypergraph regularized autoencoder(DAHDA)is proposed.First,more robust features are extracted by denoising autoencoder to improve the under-fitting problem;Then,the maximum mean discrepancy criterion is used to match the marginal and conditional distribution to solve the under-adaptation problem;Further,the hypergraph regularization term is introduced according to the relationships among source and target domain samples to solve the negative-transferproblem by discovering the multiple relationships.The loss function of the classifier is obtained according to the true labels of the source domain;Finally,the network parameters are learned by gradient descent method and the classification of the target domain samples is completed.Thirdly,for heterogeneous domain adaptation,the current methods can not fit the data distribution well and further obtain more efficient feature expression based on the shallow structure.Besides,the matching of data distribution between domains and the consistency of label and structure are not taken into account simultaneously.Aiming at this problem,a heterogeneous domain adaptation network based on autoencoder(HDANA)is proposed.First,the data of source and target domains are mapped into a shared feature space by two autoencoder networks respectively and the maximum mean discrepancy criterion is used to match the marginal and conditional distribution;Then,the manifold alignment terms are introduced,in which the geometrical term is used to keep the consistency of geometric structure,the similar and dissimilar terms are used to keep the consistency of label information between domains;Further,the loss function of the classifier is obtained according to the true labels of the source and target domains;Finally,the network parameters are learned by gradient descent method and the classification of the unlabeled target domain samples is completed.Experimental results on multiple datasets show that compared with traditional homogeneous and heterogeneous methods,the proposed methods can achieve better classification performance and solve the problem cross-domain knowledge transfer more effectively.