The Research Of Feature Adaptation Transfer Learning

In the current research boom in the field of artificial intelligence,as a major branch of machine learning,transfer learning has become a research hotspot and achieved remarkable research results.Because we are in an era of rapid growth of massive data,including various data such as voice,video,images,text,etc.,machine learning needs to use this massive data to train for optimal models in order to solve different problems to suit different application scenarios.However,there is an urgent problem behind the large amount of data,which is the lack of information annotation for big data.It takes a long time and a lot of cost to manually label data.The optimization model of transfer learning training does not need a lot of data annotation.Through the existing knowledge of the source domain of annotation data,it can help the target domain to learn,so as to solve the problem of lack of annotation data.Data classification is similar to annotation.For the application of image classification,unlabeled target domain images can be classified with the help of labeled source domain information.Therefore,transfer learning algorithms have been extensively researched and developed in the fields of data annotation and classification.Transfer learning uses the similarity between data and tasks in the source and target domains to transfer the knowledge of the source domain to the target domain,thereby training an optimization model for the target domain.Therefore,the key of the transfer learning algorithm is to measure the distance between the source domain and the target domain,and further narrow the discrepancy in data distribution between the two domains.This thesis studies the transfer learning method and applies it to image classification.Using the most important feature information of the image,on the premise of narrowing the discrepancy in feature distribution between the two domains,a better feature representation is obtained.This thesis proposes two feature adaptation transfer learning methods.1.Tensor is the generalization of vector and matrix,which is especially suitable for representing multilinear relationships that cannot be naturally represented by vector or matrix.However,most of the existing transfer learning methods are designed with the feature representation based on vector,which is difficult to represent and preserve the important structural information in many applications.Based on this,this thesis proposes a new transfer learning method based on tensor representation and adaptation regularization of feature representation.When the data distribution of the source domain and the target domain is quite different,it will distort the data too much if only the feature of the source domain are aligned with the potential shared subspace.To alleviate this problem,a method of joint domain alignment is proposed in this thesis,which aligns the data of source domain and target domain under tensor representation and aligns the shared tensor subspace simultaneously.Moreover,In order to reduce the differences in data distribution between source and target domains and to preserve the manifold consistency between samples,this thesis introduces adaptation regularization in tensor-based subspace learning,which consists of dynamic distribution alignment and graph adaptation.Finally,the fusion of joint domain alignment and adaptation regularization obtains new feature representations of the source and target domains based on the tensor subspace by joint optimization solving the shared tensor subspace.In this thesis,a large number of experiments on image classification have been conducted on several public datasets.The experimental results show that the proposed method is more robust than other mainstream transfer learning methods,which validates the proposed method.2.Most of the existing transfer learning algorithms in solving the problem of image classification often have the class imbalance problem of image.The number of samples in different class is different,which causes the data distribution to be skewed to the class with more samples,and the class with fewer samples are ignored.This will affect the transfer result of the source domain to the target domain.This thesis proposes a new transfer learning method,which is a transfer learning method based on class balance and representation learning.An random oversampling algorithm for class balance is proposed to control the class balance between source and target domains and reduce the distribution discrepancy between source and target domains.At the same time,the representation learning part based on graph structure learning and Hilbert-Schmidt independence criterion is proposed,which is introduced into the sample space of class balance for solving.Graph structure learning reduces intra-class distances,and the Hilbert-Schmidt independence criterion preserves the intrinsic dependencies of features and labels under class balance.Finally,the optimized random oversampling algorithm for class balance and the representation learning are combined to obtain the optimized domain adaptation feature representation.Through experiments on several image datasets,compared with other transfer learning methods,it effectively improves the accuracy of image classification transfer tasks.