Adversarial Domain Adaptation Method And Its Application

When there are no labeled samples or few labeled samples,the model constructed by machine learning is prone to over fitting,poor expression ability and poor generalization.For example,extreme learning machine is prone to low accuracy and poor generalization in unsupervised learning environment.So,how to use the existing labeled dataset to classify the unlabeled dataset quickly and effectively?Domain adaptation is a new method to solve this problem.Domain adaptation can realize the adaptation from rich label source domain to unlabeled target domain,which is the inevitable choice to solve the above problems.In this paper,we study domain adaptation,construct domain adaptation model to realize cross domain adaptation,design image recognition model based on transfer learning,propose two unsupervised machine learning models,and apply them to pattern recognition and medical diagnosis scenarios.The main work of this paper is as follows.Firstly,a adversarial domain adaptation extreme learning machine(ADAELM)is proposed.The model integrates extreme learning machine and adversarial learning by introducing adversarial learning.Firstly,the labeled source domain data is used to pre-train the model,and then the pre-trained model parameters are used as the initialization parameters of the adversarial domain adaptation model,and the source domain and the target domain models play games with the discriminator,so as to reduce the differences between the target domain and the source domain after they are projected into a common feature space and find the optimal target domain features.Finally,the conjugate gradient extreme learning machine(CGKELM)is used to classify the features of the optimal target domain features,and the final classification result is obtained.A large number of experiments verify the effectiveness of the method,and the effectiveness of the proposed method is verified on MNIST,USPS,OFFICE-CALTECH datasets and the medical datasets named COVID-19.At the same time,the comparative experiments with other methods show the superiority of this method.The results show that the method is fast and stable.Secondly,a unsupervised manifold adversarial domain adaptation method is proposed.Firstly,the pre-trained model of the source domain data is used as the initialization parameter of the model,and then the domain differences between the target domain and the source domain are effectively reduced through the competition between the discriminator and the feature extractor.Embedding manifold learning can effectively reduce the intra-class differences,so that the distance between the data of the same category is gradually reduced to further data alignment,this assist the process of adversarial learning and play the role of adversarial learning in a greater extent.The image recognition experiments are carried out on the public data set Office31,the effectiveness of the method is verified.The tumor diagnosis is carried out on the Break His cancer data set and the Gastric epithelial tumor data set,the experimental results show that the method has good generalization performance,and has advantages in training speed compared with deep learning.