Feature and instance joint transfer learning method in semi-supervised scene

Abstract

The invention discloses a feature and instance joint transfer learning method in a semi-supervised scene. The method aims at solving the problem of insufficient labeled data during classification model training in some fields. Data in other fields and label-free data in the field are introduced to assist training. Meanwhile, the difference of data distribution between the fields is considered, andfor the situation that a small amount of data with labels and a large amount of label-free data exist in a target domain, the invention provides a mixed balanced distribution adaptation method and aself-learning instance migration method. A feature and instance joint transfer learning method FSJT is constructed on the basis.

Description

technical field [0001] The invention relates to the technical field of transfer learning in machine learning, and specifically refers to a feature and instance joint transfer learning method in a semi-supervised scene. Background technique [0002] In 2005, the U.S. Defense Advanced Research Projects Agency (DARPA) Information Processing Technology Office officially gave a definition of transfer learning, that is, the ability to apply knowledge and skills learned in other tasks to new tasks. Compared with multi-task learning, transfer learning is more concerned with the target task, rather than learning all the source and target tasks at the same time. The roles of source and target tasks in transfer learning are no longer symmetric. The task space that needs to be studied in transfer learning is called the target domain (Target domain), and the task space related to it before is called the source domain (Source domain). The development of transfer learning at this stage m...

AI Technical Summary

Problems solved by technology

However, current transfer learning methods mainly utilize a small amount of labeled data in the target domain, or directly utilize unlabeled data

Images