Research On Image Classification Algorithm Based On Unsupervised Transfer Learning

With the advent of the big data era of intelligent information,the amount of data and data types has grown tremendously.With the continuous improvement of machine learning theory,machine learning has made great progress in real life applications.The purpose of traditional machine learning is to find the model with the least expected risk on the test data by minimizing the regularized empirical risk on the training data,but this model assumes that the training data and the test data have the same joint probability distribution.For example,in practical applications such as language recognition,natural language processing,text processing,computer vision,and data mining,the above model assumptions are basically unsatisfactory.How to take a walk based on the existence of different distributed data,quickly implement model construction,and complete data annotation is the most challenging frontier in contemporary machine learning.Domain adaptation is a representative method in transfer learning.Domain adaptive learning is an adaptive learning algorithm where the source domain data sample and the target domain data sample have the same data characteristics and categories but different data distributions.I hope to build a data model in the source domain and make it unknown in the target domain through knowledge transfer The label can be executed,so that the model assumption that the source domain and the target domain are independently and identically distributed can be ignored to improve the learning performance of the test data.At present,there are two main methods in the transfer learning method,namely the data center method and the subspace-centered method.In the data-based method,the main difference is to reduce the distribution deviation by reducing the marginal probability distribution difference and the conditional probability distribution difference at the same time.However,the two distribution differences in this type of algorithm have the same weight,resulting in poor performance in practical applications.In addition,the subspace-based method only reduces the geometric offset of the subspace,does not take into account the distribution difference between the two domains,and also does not take into account the discriminative feature learning of the source and target domain categories.Therefore,in this paper,for these two problems,a joint matching method based on feature matching and instance weighting and a joint matching method based on joint feature distribution and geometric alignment are proposed.The main work and innovations of this article are:In order to realize the domain adaptation based on the data-centric method,the important problem that the edge distribution and conditional distribution weights are not necessarily the same in different models is ignored,and a joint matching method based on feature matching and instance weighting is proposed.This method first increases the correlation between the source domain data sample and the target domain data sample by re-weighting the source domain instance,and secondly reduces the distribution offset by dynamically adjusting the distribution weight between the two domains through the balance factor,thereby achieving a balanced probability distribution Match.Simulation experiments show that the proposed framework can effectively improve the classification accuracy of transfer learning.Aiming at realizing field adaptation based on subspace method,a joint matching method based on joint feature distribution and geometric alignment is proposed.This method reduces both the subspace geometric offset and the distribution offset and makes the samples separable in the feature space.Experiments on multiple data sets show that the proposed framework can effectively improve the classification accuracy of transfer learning.