Research On Transfer Learning Method For Universal Domain Adaptation

Currently,deep learning is everywhere in our lives.A large amount of data is required for a successful deep neural network application;but in reality,we have relatively little data,and it is unrealistic to find a large amount of useful data to train us on.Therefore,in order to expand the data set,transfer learning is proposed,which can relieve the two limitations of large training data and large training cost.After that,transfer learning has become the focus of scientists’ exploration and applied in various fields.Domain adaptation is a transfer learning problem with stricter constraints,and it is also a hotspot of transfer learning research.The problem studied in this paper belongs to a subcategory of domain adaptation direction—universal domain adaptation problem.The universal domain adaptation problem proposed in 2019 means that in common life scenarios,the target domain data is unlabeled data,and it is impossible to determine the public categories and their respective private categories of the source and target domains.In this context,Universal Domain Adaptive Network is the first method proposed to solve this problem.The universal adaptation network method realizes the sample-level migration,uses two discriminators to calculate the weight of the samples,successfully marks the non-public label set as an unknown class,and then distinguishes the public class from the private class.After the universal domain adaptive network method was proposed,we found that the method ignores the complex structure in each domain,which leads to misplaced matching of data distributions between categories in the common label set.On this basis,universal adaptation network based on multiple discriminators(MUAN)is a new method proposed by this paper.The method utilizes multiple adversarial class discriminators to align between classes,measures the distribution distance between public classes in two domains through correlation alignment,and minimizes the distance loss.The method proposed in this paper can guarantee the classification accuracy of the model in the fair experimental environment and improve the transfer ability of the model.Compared with the previous method,there are mainly two major improvements.A multi-adversarial class discriminator is first used on the base framework for more fine-grained alignment of the common label set.In addition,the method proposed in this paper further enforces the narrowing of the distribution distance between the public classes of the source and target domains by using correlation alignment.Comparison with some existing popular domain adaptation methods,the MUAN method proposed in this paper not only pays attention to the multimodal structure in the common label set of each domain,prevents class alignment errors and can effectively avoid negative transfer,but also promotes the positive transfer of the model by utilizing correlation alignment.The method proposed in this paper greatly expands the capabilities of deep adversarial domain adaptation to address more transfer learning problems.In recent years,most of the papers on the universal domain adaptation problem in the field do not have a unified experimental comparison standard,but this paper compares the classification accuracy of the popular method and the multi-discriminatorbased universal domain adaptive network method trained on the target domain.Comparative experiments can fairly show the pros and cons of each method.In addition,this paper also evaluates five comparative experiments of our method in terms of algorithm performance.From the experimental results,we can know that the universal adaptation network based on multiple discriminators method proposed in this paper has better transfer effect and can effectively reduce the distance between the data distribution of public classes in the source and target domains and implement more finegrained adaptation between categories.