Deep transfer learning method of domain adaptive network

Abstract

The present invention provides a deep migration learning method for a domain-adaptive network. According to the distribution difference corresponding to each task-related layer, classification error rate and mismatch degree, the value of the loss function of the domain-adaptive network is determined, wherein any The distribution difference corresponding to the task-related layer is the distribution difference between the probability distribution of the features in any task-related layer corresponding to the source domain and the target domain respectively; and based on the value of the loss function, the parameters of the domain adaptive network are updated to Adapting the domain-adaptive network to the target domain; thereby taking the distribution difference between the probability distributions of the features in each task-related layer corresponding to the source domain and the target domain respectively as an integral part of the value of the loss function of the domain-adaptive network, Each task-related layer of the deep network is matched in different fields at the same time, and the difference between the marginal distribution and the conditional distribution in different fields is better corrected, which ensures the reliability of transfer learning and finally ensures the effect of domain-adaptive network transfer learning. .

Description

technical field [0001] The present invention relates to the technical field of computer data analysis, and more specifically, relates to a deep migration learning method of a domain-adaptive network. Background technique [0002] Internet technology has been widely used in various fields of life. Since the growth rate of unstructured data such as text, images, and videos is increasing rapidly, it is necessary to propose analysis methods and processing algorithms for these data. Large-scale unstructured data can be collected through various information channels, but most of the data lack information such as labels, which means that conventional supervised learning is difficult to apply to these data. [0003] In order to cope with the scarcity of labeled data, a semantic network knowledge base based on group wisdom was launched. With the help of Internet users, a nearly unlimited knowledge resource, it can label and maintain large-scale data in some important fields, such as ...

AI Technical Summary

Problems solved by technology

Domain Adaptation Network (Domain Adaptation Network) is a deep neural network. The model learned from the original problem (domain) can be well adapted to a different target problem (domain). Domain Adaptation Network often includes multiple layers structure, while the transferability of features is significantly weakened in the middle layer of the domain-adaptive network, and severely reduced in the upper layer of the domain-adaptive network; in recent years, the main challenge of transfer learning is the reliability of the transfer learning process, that is, it is difficult to guarantee the domain The transfer effect of the features of each layer in the adaptive network, especially the transfer effect of the upper layer features, thus affecting the effect of domain adaptive network transfer learning

Images