Research Of Unsupervised Domain Adaptation Algorithms Based On Deep Convolutional Neural Networks

Deep neural networks have achieved extensive success in both theory and practice in recent years.However,there are two constraints during the training process.The first one is that a large number of labeled data is necessary to prevent the model from over fitting.The second is that the training and testing data must be sampled independently from an identical probability distribution.These two conditions seriously restrict the possibility of applying deep learning model to practical problems.A popular method to overcome these two difficulties is unsupervised domain adaptation,which aims to solve the classification problem of the target domain with different data distribution by using the existing labeled data from source domain.This paper designs unsupervised domain adaptation algorithms based on deep neural networks,using deep neural networks to extract domain invariant deep feature representation in the source domain to solve the image classification task in the target domain.It mainly focuses on the following two contents:Firstly,a deep convolutional neural network architecture is proposed based on the spatial pyramid pooling layer and the gating mechanism,and its corresponding unsupervised domain adaptation algorithm is presented.In this model,pyramid pooling layer is used to extract features from different levels of the feature maps,and then a gating mechanism is used to save the task related knowledge of each sample in the classification task as the global task knowledge vector,and the global task knowledge vector is used to recalibrate the features of each sample.In the unsupervised domain adaptation algorithm of this model,samples in different domains use different encoders to extract features,while using the same global task knowledge vector to recalibrate the sample features from different domains,which realizes knowledge transfer and reuse.Secondly,the maximum entropy loss based unsupervised domain adaptation method is proposed to achieve the conditional probability distribution alignment of features between the source and target domains.Motived by the idea of clustering methods,the weight vector of the classifier is regarded as the center of each category cluster,and the classification feature distribution alignment is realized by alternately updating the classifier and encoder.In the classifier updating step,the position of the weight vector in the classifier is adjusted by maximizing the entropy of the sample prediction probability of the neural network in the target domain,so that the weight vector of the classifier moves to the cluster center of the target domain;in the encoder updating step,the feature distribution in the source domain is adjusted by minimizing the classification loss in the source domain,to guarantee that the feature in the source domain is moved to the weight vector of the classifier.In order to align the marginal feature distribution from different domains,the maximum entropy loss based unsupervised domain adaptation method also combines the idea of gradient reversal layer in the domain adversarial neural network,which trains both the encoder and the domain discriminator in a single gradient descent step.