The Research Of Multi-Task Learning From Multiple Sources

Traditional learning techniques have the assumption that training and test data are drawn from the same data distribution,however the assumption in the real-world is not true.The traditional learning algorithm not only requires high quality data and needs to reduce noise but also needs the artifacts to extract the characteristics of the data.When the sample size is very large,it needs to consume a lot of Manpower.In the era of big data,a remaining challenge is diversity of data from different sources and how to apply the knowledge learned from multiple source domains to a target domain.In general,the data of different source domains have interclass and intraclass correlation,and the distribution of these source domain data is different,and the correlation between classes and classes in the fusion domains and their distribution differences are established.The classification model has important significance.This paper proposed a consensus regularization framework based on multi-task depth learning,which is suitable for migration learning from multiple source domains to target domains.In this framework,multidimensional depth learning network training is used for each source domain.Each source domain learns multiple tasks in parallel,improves the efficiency of the algorithm,and establishes the correlation of multiple tasks.The tasks are interdependent.The algorithm has better classification performance.The results of the source domain training model not only consider the data characteristics in this source domain,but also consider the model of the other source domain training.The prediction of the data is consistent with the prediction of the source domain training model.In addition,this paper provides a theoretical analysis of the consensus regularization framework based on depth learning,and applies this algorithm to the field of image classification.By analyzing the experimental results,the validity of this learning method is verified.